Research proposals 40th cycle

Ultra-low latency multimedia streaming over HTTP/3

Media Quality Optimization using Machine Learning on Large Scale Datasets

In recent years, ML has been successfully employed to develop video quality estimation algorithm (see, e.g., the Netflix VMAF proposal) to be integrated in media quality optimization frameworks. However, despite these improvements, no technique can currently be considered reliable, partly because the inner workings of ML models cannot be easily and fully understood especially when they are based on "black box" neural network models.

We aim to improve the situation by developing more reliable and explainable quality prediction models. Starting from Internet Media Group's ongoing work on modeling the behavior of single human subjects in media quality experiments, the candidate will derive a systematic approach by employing several subjectively annotated datasets (i.e., with quality scores given by human subjects). With such an approach we expect to be able to identify meaningful media quality features useful to develop new reliable and explainable quality prediction models.
However, to identify and improve such features by using machine learning models, it is important to include also large-scale, not subjectively annotated, datasets. To efficiently deal with this large amount of data, it is necessary to develop a framework comprising a set of tools that allows to more easily process both the subjective scores (given by human subjects) as well as objective scores in an efficient and integrated manner, since currently every dataset has its own characteristics, quality scale, way of describing distortions, etc. which make integration difficult. Such framework, that we will make publicly available for research purposes, will constitute the basis for reproducible research, which is increasingly important for ML techniques.
The framework will allow to systematically investigate existing quality prediction algorithms finding strength and weaknesses, as well as to identify the most challenging content on which newer development can be based.
Such objectives will be achieved by using both theoretical and practical approaches. The resulting insight will then be validated in practical cases by analyzing the performance of the system with simulations and experiments with industry-grade signals, leveraging cooperation with companies to facilitate the migration of the developed algorithms and technologies into prototypes that can then be effectively tested in real industrial media processing pipelines.

The workplan of the activities is detailed in the following. In the first year the PhD candidate will first familiarize with the recently proposed ML and AI-based techniques for media quality optimization, as well as the characteristics of the publicly available datasets for research purposes. 
In parallel, a framework will be created to efficiently process the large sets of data (especially for the video case) with potentially complex measures that might need retraining, fine-tuning or other computationally complex optimizations. It is expected to make this framework publicly available also to address the research reproducibility issues that are of growing interest in the ML community. This initial investigation and activities are expected to lead to conference publications.
In the second year, building on the framework and the theoretical knowledge already present in the research group, new media quality indicators for specific quality features will be developed, simulated, and tested to demonstrate their performance and in particular their ability to identify the root causes of the quality scores for several existing quality prediction algorithms, thus partly explaining their inner working methods in a more understandable form. In this context, potential shortcomings of such algorithms will be systematically identified. These results are expected to yield one or more journal publications.
In the third year the activity will then be expanded to propose improvements that can mitigate the identified shortcoming as well as to create proposals for quality prediction algorithms based on the previously identified robust features. Such proposal will target journal publications.

Possible targets for research publications, well known to the proposer, include IEEE Transactions on Multimedia, Elsevier Signal Processing: Image Communication, ACM Transactions on Multimedia Computing Communications and Applications, Elsevier Multimedia Tools and Applications, various IEEE/ACM international conferences (IEEE ICME, IEEE MMSP, QoMEX, ACM MM, ACM MMSys).

The proposer is actively collaborating with the Video Quality Experts Group (VQEG), an international group of experts from academia and industry that aims to develop new standards in the context of video quality. In particular the tutor is co-chair of the JEG-Hybrid project which is very interested in the activity previously described.

Security Analysis and Automation in Smart Systems

Local energy markets in citizen-centered energy communities

Simulation and Modelling of V2X connectivity with traffic simulation

Machine Learning techniques for real-time State-of-Health estimation of Electric Vehicles batteries

Multi-Device Programming for Artistic Expression

Media artists use smartphones and IoT devices as material for creative exploration. However, some do not code and have a low interest in learning. In addition, programming artworks enclose characteristics that differ from traditional coding.

This Ph.D. proposal aims to extend our comprehension of the needs of artists for creative coding through the design, implementation, and evaluation of toolkits that serve them to realize code-based artworks across multiple devices and media effectively.

The recent availability of smartphones, AR/VR headsets, IoT-enabled devices, and microcontroller kits creates new opportunities for creative explorations for media artists and designers. The field of creative coding emphasizes the goal of expression rather than function, and creative coders combine computational skills with creative insight. In some cases, artists and designers are interested in creative coding but lack the knowledge or programming skills to benefit from the offered possibilities.

The main research objective of this Ph.D. proposal is to extend our comprehension of the needs of media artists and designers for creative coding across multiple devices and media. To reach this objective, the Ph.D. student will study, design, develop, and evaluate proper models and novel technical solutions (e.g., toolkits and tools) for supporting creative coders. The proposal envisions focusing on both creative coders and end-user programmers. The work will start from the needs of the stakeholders (i.e., artists and designers), complemented by the existing literature. Using a participatory approach, the Ph.D. student will keep the stakeholders involved in the various phases of the work.

In particular, the Ph.D. research activity will focus on the following: - Study of the creative coding field, stakeholders' needs and current tools, and HCI techniques able to support the identification of suitable requirements and the creation of technical solutions to effectively support creative exploration and coding. - Creation of a theoretical framework to satisfy the identified needs and requirements, able to adapt to different media, devices, and skills. For instance, it can include end-user personalization as a way to allow end-users to create code-based artifacts and AI techniques to support the creation of programs. - Development of a toolkit and related tools to experiment with the theoretical framework's facets. The creation and evaluation of the tools will serve as the validation for the framework.

The work plan will be organized according to the following four phases, partially overlapped:- Phase 1 (months 0-6): literature review about creative coders and coding; focus groups and interviews with designers and media artists of various skills; definitions and development of a set of use cases and promising strategies to be adopted.- Phase 2 (months 6-18): research, definition, and experimentation of an initial version of the theoretical framework and a first toolkit for creative coders, starting from the outcome of the previous phase. In this phase, the focus will be on the most common target devices, i.e., the smartphone and the PC, with the design, implementation, and evaluation of suitable tools.- Phase 3 (months 12-24): research, definition, and experimentation of a second toolkit (or an evolution of the previous one) for novice creative coders and end-users. Such a toolkit uses artificial intelligence and machine learning to help during the coding process. The subsequent design, implementation, evaluation of suitable tools, and extension of the framework.- Phase 4 (months 24-36): extension and generalization of the previous phases to include additional target devices and consolidate the theoretical framework; evaluation of the toolkit and the tools in real settings with a large number of artists.

It is expected that the results of this research will be published in some of the top conferences in the Human-Computer Interaction field (e.g., ACM CHI, ACM CSCW, ACM C&C, and ACM IUI). One or more journal publications are expected on a subset of the following international journals: ACM Transactions on Computer-Human Interaction, ACM Transactions on Interactive Intelligent Systems, and International Journal of Human Computer Studies.

Digital Wellbeing By Desing

Management Solutions for Autonomous Networks

Two research questions (RQ) guide the proposed work: 

RQ1: How can we design and implement on local and larger-scale testbeds effective transport and routing network protocols that integrate the network stack at different scopes using recent advances in supervised and unsupervised learning? 

RQ2: To scale the use of machine learning-based solutions in network management, what are the most efficient distributed machine learning architectures that can be implemented at the network edge layer? 

The final target of the research work is to answer these questions, also by evaluating the proposed solutions on small-scale network emulators or large-scale virtual network testbeds, using a few applications, including virtual and augmented reality, precision agriculture, or haptic wearables. In essence, the main goals are to provide innovation in network monitoring, network adaptation, and network resilience, using centralized and distributed learning integrated with edge computing infrastructures. Both vertical and horizontal integration will be considered. By vertical integration, we mean considering learning problems that integrate states across network hardware and software, as well as states across the network stack across different scopes. For example, the candidate will design data-driven algorithms for congestion control problems to address the tussle between in-network and end-to-end congestion notifications. By horizontal learning, we mean using states from local (e.g., physical layer) and wide area (e.g., transport layer) as input for the learning-based algorithms. The data needed by these algorithms are carried to the learning actor by means of newly defined in-band network telemetry mechanisms.  Aside from supporting resiliency with the vertical integration, solutions must offer resiliency across a wide (horizontal) range of network operations:  from close-edge, i.e., near the device, to the far-edge, with the design of secure data-centric resource allocation (federated) algorithms.

The research activity will be organized in three phases:

Phase 1 (1st year): the candidate will analyze the state-of-the-art solutions for network management, with particular emphasis on knowledge-based network automation techniques. The candidate will then define detailed guidelines for the development of architectures and protocols that are suitable for automatic operation and configuration of NextG networks, with particular reference to edge infrastructures. Specific use-cases will also be defined during this phase (e.g., in virtual reality). Such use cases will help identifying ad-hoc requirements and will include peculiarities of specific environments. With these use cases in mind, the candidate will also design and implement novel solutions to deal with the partial availability of data within distributed edge infrastructures. Results of this work will likely result in conference publications. 

Phase 2 (2nd year): the candidate will consolidate the approaches proposed in the previous year, focusing on the design and implementation of mechanisms for vertical and horizontal integration of supervised and unsupervised learning with network virtualization. Network, and computational resources will be considered for the definition of proper allocation algorithms. All solutions will be implemented and tested. Results will be published, targeting at least one journal publication.

Phase 3 (3rd year): the consolidation and the experimentation of the proposed approach will be completed. Particular emphasis will be given to the identified use cases, properly tuning the developed solutions to real scenarios. Major importance will be given to the quality offered to the service, with specific emphasis on the minimization of latencies in order to enable a real-time network automation for critical environments (e.g., telehealth systems, precision agriculture, or haptic wearables). Further conference and journal publications are expected.

The research activity is in collaboration with Saint Louis University, MO, USA, also in the context of the NSF grant #2201536 ?Integration-Small: A Software-Defined Edge Infrastructure Testbed for Full-stack Data-Driven Wireless Network Applications?. Furthermore, it is related to active collaborations with Futurewei Inc. and Tiesse SpA, both interested in the covered topics.

The contributions produced by the proposed research can be published in conferences and journals belonging to the areas of networking and machine learning (e.g. IEEE INFOCOM, ICML, ACM/IEEE Transactions on Networking, or IEEE Transactions on Network and Service Management) and cloud/fog computing (e.g. IEEE/ACM SEC, IEEE ICFEC, IEEE Transactions on Cloud Computing), as well as in publications related to the specific areas that could benefit from the proposed solutions (e.g., IEEE Transactions on Industrial Informatics, IEEE Transactions on Vehicular Technology)

Preserving privacy and fairness with generative AI-based synthetic data production

Synthetic data generation enables the reproduction, diversification, and augmentation of real data in contexts where data is scarce and where preserving privacy is paramount. However, synthetic data may come at costs of unrealistic synthetic populations or limited accuracy in the downstream predictions and classifications. In addition, reliable techniques for reaching satisfactory trade-offs between contrasting requirements (e.g., privacy, fairness, and accuracy) are still object of research and experimentation, as well as how to produce suitable dataset and model documentation.
 
The goal of this PhD proposal is to design, develop and test new generative models for synthetic data production able to preserve privacy and guarantee fairness, while maintaining acceptable levels of accuracy. The focus will be mostly on tabular data because this type of data is used in most of the applications where fairness and privacy are paramount (for example, allocating social benefits or economic resources).
 
The new developed techniques will be compared with state-of-art generative models (e.g., language models, variational autoencoder, generative adversarial network, diffusion models, self-supervised learning, etc. ) and with traditional probabilistic methods for dataset generation (e.g, Bayesan networks, univariate kernel density estimation, etc.) on a variety of evaluation measures, such as: distance from original population; differential privacy; imbalance; fairness and accuracy of models trained and tested on generated data.
 
In addition, given the rising importance of auditable algorithms in the European legislative context, two further aspects will be investigated: i) how to properly document synthetic datasets and the model that generated them ; ii) how to generate suitable synthetic data for auditing black block systems against discrimination.
 
Considering the different (and mostly contrasting) dimensions of evaluation above mentioned, the high-level research questions are:
 
RQ 1. How to improve state of art generative models for synthetic data generation in a way that they can preserve privacy and fairness while maintaining acceptable levels of accuracy?  
 
RQ 2. What is the trade-off (between privacy, fairness, and accuracy) reached by the newly developed techniques in comparison of:
-      1.1 state of art generative-AI models?
-      1.2 traditional probabilistic methods?
 
RQ 3. Is it possible to facilitate auditing of black box systems
-      1.1 by documenting synthetic datasets and their generative models?
-      1.2 by producing exhaustive and realistic synthetic data for testing black box systems against discrimination?
 
Datasets used in the fair machine learning community will be firstly used as starting point, with possible integrations from incoming projects and/or industrial collaborations.
 
A possible workplan is the following.
 
Year 1:
RQ 1. - Task 1.1) Analysis of state of the art about: fairness datasets, measures for evaluation (privacy, balance, fairness, distance from real population, etc), synthetic data generation (with traditional probabilistic methods and generative AI).
 
RQ 1. - Task 1.2) Design and develop new generative models, based on existing ones.
 
RQ 2. - Task 2.1) Design and prepare experimentation according to the measures and datasets selected in task 1.1.
 
RQ 2. - Task 2.2) Run experiments, collect data and analyze results.
 
Year 2:
RQ 2. - Task 2.3) New experiments based on results of Task 2.2.

RQ 3. Task 3.1) State of art of existing documentation suites for datasets (e.g., datasheets) and models (e.g., model cards); collect and organize online software systems that are suitable candidates for black box testing against discrimination (e.g., insurance, online advertising, etc.).
 
RQ 3. - Task 3.2) Case study identification: select system for black box auditing, and produce suitable synthetic data sets with the model(s) developed in previous RQ.
 
Year 3:
RQ 3. - Task 3.3) Development of a new documentation suite for documenting synthetic datasets and their generative models, and application to the case study, including evaluation with users (to be identified).

RQ 3. - Task 3.4) Quantitative analysis of discrimination in selected existing software systems using the synthetic data generated.
 
Dissemination of results is a cross-cutting activity. Possible venues for publication are:
 
Journals:
-      IEEE Transactions on Software Engineering
-      ACM Transactions on Software Engineering and Methodology
-      Journal of Machine Learning Research
-      Empirical Software Engineering
-      ACM Transactions on Information Systems
-      European Journal of Information Systems
-      Journal of Systems and Software
-      Software X
-      ACM Journal of Responsible Computing
 
Conferences:
-      ACM Conference on Fairness, Accountability, and Transparency
-      AAAI/ACM Conference on AI, Ethics, and Society
-      International Conference on Internet Technologies & Society
-      EPIA Conference on Artificial Intelligence
-      ACM/IEEE International Symposium on Empirical Software Engineering and Measurement
-      International Conference on Frontiers of Artificial Intelligence, Ethics, and Multidisciplinary Application
-      International Conference on Software Engineering
-      Conference on Neural Information Processing Systems (NeurIPS)
 
Workshops:
-      International Workshop on Data science for equality, inclusion and well-being challenges
-      International Workshop on Equitable Data & Technology
-      Workshop on Bias and Fairness in AI

Digital Twin development for the enhancement of manufacturing systems

State-of-Health diagnostic framework towards battery digital twins

Modeling, simualtion and validation of modern electronic systems

Robust AI systems for data-limited applications

Artificial Intelligence applications for advanced manufacturing systems

AI for Secured Networks: Language Models for Automated  Security Log Analysis

Research objectives:
 
Investigate and evaluate the capabilities of LLM agents in automating the manual investigations of the security analyst. This would assist them in analysis and incident reporting.  

The candidate will perform research to determine whether, and to what extent, the recent advances in language models could be used to automate and assist security analysts in the process (i) of learning the security-device rules by example and (ii) autonomously investigating the challenging cases currently analyzed by humans.
 
In the second phase, the candidate will investigate how and if lightweight and generalizable language models can extract insights from raw data, as today large language models can do.  The goal is to investigate whether models with limited supervision and a minimal number of trusted labels can attain comparable performance to generic large language models (LLMs) when applied to specific tasks such as code understanding, classification, anomaly detection, bug detection, or identifying security breaches.
 
The research will consider multi-modal embeddings to conceptually constrain the embeddings towards the right task. 
By forcing the model to create multi-modal embeddings conceptually constrained to the right task, the model will possess the ability to generalize and autonomously reason about novel and previously unencountered tasks. For instance, test joint learning of (i) natural language label explanation of the security threat and (ii) the packet payload, using, e.g., contrastive learning techniques. 
 
The project will involve a collaboration with Huawei Technologies France and Politecnico di Torino. 

Outline of the research work plan:

1st  year- Study of the state-of-the-art of security log analysis and state-of-the-art language models in ML.- Data collection and analysis of raw and structured data on security devices such as Firewall/Intrusion Prevention Systems (IPS), Endpoint Detection and Response (EDR) and Cloud security services. 

2nd  year- Adaptation and extension solutions to learn the security-device rules by example and autonomously investigate complex cases.- Propose and develop innovative solutions to the problems of cyber threats analysis with Language models.- Propose multi-modal embeddings for network raw data and security logs.

3rd  year - Tune the developed techniques and highlight possible strategies to counteract the various threats.- Application of the strategies to new data for validation and testing.

References:
- Boffa, M., Valentim, R. V., Vassio, L., Giordano, D., Drago, I., Mellia, M., & Houidi, Z. B. (2023). LogPr\'ecis: Unleashing Language Models for Automated Shell Log Analysis. arXiv preprint arXiv:2307.08309- Boffa, M., Milan, G., Vassio, L., Drago, I., Mellia, M., & Houidi, Z. B. (2022, June). Towards nlp-based processing of honeypot logs. In 2022 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) (pp. 314-321). IEEE.- Boffa, M., Vassio, L., Mellia, M., Drago, I., Milan, G., Houidi, Z. B., & Rossi, D. (2022, December). On using pretext tasks to learn representations from network logs. In Proceedings of the 1st International Workshop on Native Network Intelligence (pp. 21-26).

List of possible venues for publications:
- Security venues: IEEE Symposium on Security and Privacy, IEEE Transactions on Information Forensics and Security, ACM Symposium on Computer and Communications Security (CCS), USENIX Security Symposium, IEEE Security & Privacy;
- AI venues: Neural Information Processing Systems (NeurIPS), International Conference on Learning Representations (ICLR), International Conference on Machine Learning (ICML), AAAI Conference on Artificial Intelligence, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD);
- Computer networks venues: Distributed System Security Symposium (NDSS), Privacy Enhancing Technologies Symposium, The Web Conference (formerly International World Wide Web Conference WWW), ACM International Conference on Emerging Networking EXperiments and Technologies (CoNEXT), USENIX Symposium on Networked Systems Design and Implementation (NSDI).

- Good programming skills (such as Python, Torch, Spark)

- Excellent Machine Learning knowledge

- Knowledge NLP and LM

- Basics of Networking and security

Leveraging Machine Learning Analytics for Intelligent Transport Systems Optimization in Smart Cities

Research objectives:
 
This PhD research seeks to harness the power of machine learning and big data analytics in understanding and optimizing mobility through the analysis of data collected from black boxes in fleets of cars. 

This proposal outlines a comprehensive plan to leverage big data analytics for intelligent transport systems in smart cities. The impact of mobility based on electric vehicles and its comparison with previous habits will be a core part of the study. 
The research objectives aim to contribute valuable insights to mobility planning and optimization, and the work plan ensures a systematic approach to achieving these objectives. 
The PhD student will be involved in research activities with companies and funded research projects. Data will be provided by a company.

Outline of the research work plan:

1st  year- Study of the state-of-the-art data analysis techniques for transportation and mobility.- Data collection, Exploration and Pre-processing: Extract and pre-process raw data from black boxes, ensuring data quality and compatibility for further analysis. Develop techniques to handle missing or incomplete data.- Investigate and implement privacy-preserving methods to ensure ethical use of mobility data while still deriving valuable insights.

2nd  year- Apply machine learning algorithms to identify patterns in mobility data, extracting insights into traffic flows, congestion, and usage patterns.- Implement anomaly detection mechanisms to identify unusual events and improve system resilience- Develop predictive models to forecast traffic conditions, enabling proactive measures to alleviate congestion and enhance overall traffic management.- Explore adaptive algorithms for real-time adjustments based on dynamic traffic patterns.

3rd  year - Integrate developed algorithms into a cohesive system for intelligent transport systems.- Validate the system using real-world scenarios and fine-tune algorithms for optimal performance.

References:- Ciociola, A., Cocca, M., Giordano, D., Mellia, M., Morichetta, A., Putina, A., & Salutari, F. (2017, August). UMAP: Urban mobility analysis platform to harvest car-sharing data. In SmartWorld/(pp. 1-8). IEEE.- Cocca, M., Giordano, D., Mellia, M., & Vassio, L. (2019). Free-floating electric car sharing: A data-driven approach for system design. IEEE Transactions on Intelligent Transportation Systems, 20(12), 4691-4703.- Cocca, M., Giordano, D., Mellia, M., & Vassio, L. (2019). Free-floating electric car sharing design: Data-driven optimisation. Pervasive and Mobile Computing, 55, 59-75.

List of possible venues for publications
- IEEE Transactions on Intelligent Transportation Systems- Elsevier Transportation Research- IEEE International Conference on Data Science and Advanced Analytics- IEEE International Conference on Big Data  - IEEE International Smart Cities Conference- ACM Transactions on Spatial Algorithms and Systems- IEEE Transactions on Vehicular Technology- Elsevier Cities

Natural Language Processing e Large Language Models for source code generation

The integration of Artificial Intelligence, especially Machine/Deep Learning, in industrial processes promises swift changes. Companies stand to benefit in the short term with improved production quality, efficiency, and automated routine tasks, fostering positive impacts on work environments. In addition to Natural Language Processing, Large Language Models (LLMs) have already demonstrated significant progress in healthcare, education, software development, finance, journalism, scientific research, and customer support. The future entails optimizing LLMs for widespread use, enhancing the competitiveness of the industrial system and streamlining collaborative supply chain management.

The objective of this Ph.D. proposal consists of the design and development of AI-assisted models based on Natural Language Processing (NLP) and Large Language Models (LLMs) to optimize the AI-assisted source code generation in the context of software development by enhancing the process, leading to more efficient, expressive, and context-aware.

During the three years of the Ph.D., the research activity will be divided into five phases:- Survey existing literature on NLP applications in software engineering and analyze methodologies and challenges in source code generation using language models.- Design and develop Large Language Models for improved programming language understanding by investigating techniques for domain-specific customization of language models.- Develop algorithms and strategies for context-aware source code generation by implementing prototype systems for evaluation and refinement.- Design and implement a collaborative framework that seamlessly integrates developer input with language model suggestions.- Evaluate the effectiveness of the collaboration framework through user studies and real-world projects.

Possible international scientific journals and conference:- IEEE Transactions on Audio, Speech, and Language Processing- IEEE Transactions on Software Engineering- IEEE Transaction on Industrial Informatics,- IEEE Transactions on Industry Applications,- Engineering Applications of Artificial Intelligence,- Expert Systems with Applications,- IEEE NLP-KE internat. conf.- IEEE ICNLP internat. conf.- IEEE Compsac internat. conf.

Cloud continuum machine learning

Graph network models for Data Science

Automatic composability of Large Co-simulation Scenarios for smart energy communities

A complex system such as a multi-energy system requires the accurate modelling of the heterogeneous aspects that constitute the overall phenomena under study. To achieve this goal researchers in different fields have started using co-simulation and model coupling to build new models capable of describing the interactions and the overall complexity. Such approaches give the possibility of coupling different models, running on different simulators and/or simulation engines, by exchanging data via some standard protocols over the internet. Indeed, such models have been developed and validated following a methodology that can be compared to service-oriented architecture, thus, reducing the time and complexity of building new models from scratch. Moreover, such an approach disease the interconnection of the vertical knowledge coming from each discipline/domain that is involved in the complex system, eg. ICT or Energy experts. Examples of models can be software entities that replicate the realistic behaviour of a photovoltaic (PV) system, energy storage, heating distribution networks or, even, human beans. Nowadays, researchers have invested in the usage of co-simulation orchestrators to achieve the goal of interconnection and synchronization of different models and simulators, including real-time simulators. However, the setup of the co-simulation is not an easy and trivial task as it is time-consuming and it requires the involvement of domain and co-simulation experts. This research topic aims to develop a framework, that exploits existing co-simulation orchestrators, for the automatic composability of co-simulation scenarios in a distributed infrastructure to assess different aspects of Multi-Energy-Systems. The framework will integrate models in a plug-and-play fashion reducing as much as possible the coding phase and the presence of a co-simulation expert easing the work of multi-energy systems engineers. Moreover, the framework will ease the setup in terms of computational resources for the modelling of complex and large scenarios. The final purpose consists of simulating the impact and management of future energy systems to foster the energy transition. Thus, the resulting infrastructure will integrate with a semantic approach in a distributed environment heterogeneous i) data sources, ii) cyber-physical-systems, i.e. Internet-of-Things devices iii) models of energy systems and iv) real-time simulators. The starting point of this activity will be the already existing EC-L co-simulation platform, which will be enhanced by embedding all the aforementioned features.

Hence the research will focus on developing:- a methodology based on semantic web technologies for linking and interconnecting simulators automatically in a co-simulation approach- a domain-specific ontology for describing the components and interconnection of multi-energy system models- a methodology for the automatic composability and setup of the distributed infrastructures of the energetic scenario to assess (e.g., the impact of PV systems and EVs in a city)
In a nutshell, the final result will provide a tool that exploits visual programming, semantic representation and cloud technologies to offer co-simulation as a service to describe multi-energy systems simulation scenarios in a plug-and-play fashion opening the usage of cosimulation to a wider audience.

The outcomes of this research will be a distributed co-simulation platform for:- planning the evolution of the future smart multi-energy system by taking into account the operational phase- evaluating the effect of different policies and related customer satisfaction- evaluating the performances of hardware components in a realistic test bench

During the first year, the candidate will study the literature solutions of existing co-simulation platforms to identify the best available solution for i) large-scale smart energy system simulation in distributed environments and ii) semantic web solutions to describe complex systems with a focus on the multi-energy system domain. Finally, the student will design the overall framework starting from the requirements identification and definition. 

During the second year, the candidate will face the implementation of the visual and semantic framework for model coupling and scenario creation. Furthermore, the candidate will start developing software solutions to automatic composability and setup of the co-simulation environments in terms of simulator deployment in a cloud system. 

During the third year, the candidate will complete the overall framework development and test it in different case study scenarios to assess the capabilities of the platform in terms of automatic scenario composition and setup.     

Possible international scientific journals and conferences:- IEEE Transaction Smart Grid- IEEE Transaction on Industrial Informatics,- IEEE Transaction on sustainable computing,- IEEE EEEIC internat. conf.- IEEE SEST internat. conf.- IEEE Compsac internat. conf.

Multivariate time series representation learning for vehicle telematics data analysis

Designing a cloud-based heterogeneous prototyping platform for the development of fog computing apps

Designing a Development Framework for Engineering Edge-Based AIoT Sensor Solutions

Computational Intelligence for Computer-Aided Design

The recent fortune of the term "Machine Learning" renewed the interests in many automatic processes; moreover, the publicized successes of (deep) neural networks smoothed down the bias against other non-explicable black-box approaches, such as Evolutionary Algorithms, or the use of complex kernels in linear models. The goal of the research is twofold: from an academic point of view, tweaking existing methodologies, as well as developing new ones, specifically able to tackle CAD problems; from an industrial point of view, creating a highly qualified expert able to bring the scientific know-how into a company, while being also able to understand the practical needs, such as how data are selected and possibly collected. The need to team the experts from industry with more mathematically minded researchers is apparent: frequently a great knowledge of the practicalities is not accompanied by an adequate understanding of the statistical models used for analysis and predictions.

In the first year, the research will consider techniques less able to process large amount of information, but perhaps more able to exploit all problem-specific knowledge available. It will almost certainly include bio-inspired techniques for generating, optimizing, minimizing test programs; statistical methods for analyzing and predicting the outcome of industrial processes (e.g., predicting the maximum operating frequency of a programmable unit based on the frequencies measured by some ring oscillators; detecting dangerous elements in a circuit; predicting catastrophic events). The activity is also like to exploit (deep) neural networks, however developing novel, creative results in this area is not a priority. On the contrary, the research shall face problems related to dimensionality reduction, feature extraction and prototypes identification/creation.

Then the research shall also focus on the study of surrogate measures, that is, the use of measures that can be easily and inexpensively gathered as a proxy for others, more industrially relevant but expensive. In this regard, the tutors are working with a semiconductor manufacturer for using in-situ sensors values as a proxy for the prediction of operating frequency. The work could then proceed by tackling problems related to "dimensionality reduction", useful to limit the number of input data of the model, and "feature selection", essential when each single feature is the result of a costly measurement. At the same time, the research is likely to help the introduction of more advanced optimization techniques in everyday tasks.

From a practical standpoint, starting in the second year, the activity would continue by analyzing a current practical need, namely: "predictive maintenance". A significant amount of data is currently collected by many industries, although in a rather disorganized way. The student would start by analyzing the practical problems of data collection, storage, and transmission, while, at the same time, practicing with the principles of data profiling, classification, and regression (all topics that are currently considered part of "machine learning"). The analysis of sequences to predict the final event, or rather identify a trigger, is an open research topic, with implications far beyond CAD. Unfortunately, unlikely popular ML scenarios, the availability of data is a significant limitation, a situation where the amount of available data for training is insufficient and is sometimes labeled "small data". 

Expected target publications:

Top journals with impact factors

* ASOC -- Applied Soft Computing
* TEC -- IEEE Transactions on Evolutionary Computation
* TC -- IEEE Transactions on Computers

Top conferences

* ITC -- International Test Conference
* DATE -- Design, Automation and Test in Europe Conference
* GECCO -- Genetic and Evolutionary Computation Conference
* CEC/WCCI -- World Congress on Computational Intelligence
* PPSN - Parallel Problem Solving From Nature

Notes:

* The CAD Group has a long record of successful applications of intelligent systems in several different domains. For the specific activities, the list of possibly involved companies include: SPEA, Infineon (through the Ph.D. student Niccol? Bellarmino), ST Microelectronics, Comau (through the Ph.D. student Eliana Giovannitti)
* The tutors are collaborating with Infineon on the subjects listed in the proposal: Two contracts have been signed, the third extension is currently under discussion; A joint paper has been published at ITC, other one was submitted, others are in preparation.
* The tutors are collaborating with SPEA under the umbrella contract "Colibri". Such a contract is likely to be renewed on precisely the topics listed in the proposal.

Security of Software Networks

Emerging Topics in Evolutionary Computation: Diversity Promotion and Graph-GP

Although the classical approach to representing solutions in EC involves bit strings and expression trees, far more complex encodings have been recetly proposed. More specifically, graph-based representations have led to novel applications of EC in circuit design, cryptography, image analysis, and other fields.

At the same time, divergence of character, or, more precisely, the lack of it, is widely recognized as the most impairing single problem in the field of EC. While divergence of character is a cornerstone of natural evolution, in EC all candidate solutions eventually crowd the very same areas in the search space, such a "lack of speciation" has been pointed out in the seminal work of Holland back in 1975. It is usually labeled with the oxymoron "premature convergence" to stress the tendency of an algorithm to convergence toward a point where it was not supposed to converge to in the first place. The research activity would tackle "diversity promotion", that is either "increasing" or "preserving" diversity in an EC population, both from a practical and theoretical point of view. It will also include the related problems of defining and measuring diversity.

The research project shall include an extensive experimental study of existing diversity preservation methods across various global optimization problems. Open-source, general-purpose EA toolkits, inspyred and DEAP, will also be used to study the influence of various methodologies and modifications on the population dynamics. Solutions that do not require the analysis of the internal structure of the individual (e.g., Cellular EAs, Deterministic Crowding, Hierarchical Fair Competition, Island Models, or Segregation) shall be considered. This study should allow the development of a, possibly new, effective methodology, able to generalize and coalesce most of the cited techniques.

During the first year, the candidate will take a course in Artificial Intelligence, and all Ph.D. courses of the educational path on Data Science. Additionally, the candidate is required to improve the knowledge of Python.

Starting from the second year, the research activity shall include Turing-complete program generation. The candidate will work on an open-source Python project, currently under active development. The candidate will try to replicate the work of the first year on much more difficult genotype-level methodologies, such as Clearing, Diversifiers, Fitness Sharing, Restricted Tournament Selection, Sequential Niching, Standard Crowding, Tarpeian Method, and Two-level Diversity Selection.

At some point, probably toward the end of the second year, the new methodologies will be integrated into the Grammatical Evolution framework developed at the Machine Learning Lab of University of Trieste ? GE allows a sharp distinction between phenotype, genotype and fitness, creating an unprecedented test bench (the research group is already collaborating with a group in UniTS on these topics, see "Multi-level diversity promotion strategies for Grammar-guided Genetic Programming" Applied Soft Computing, 2019).

A remarkable goal of this research would be to link phenotype-level methodologies to genotype measures.

Target Publications

Journals with impact factors
- ASOC - Applied Soft Computing
- ECJ - Evolutionary Computation Journal
- GPem - Genetic Programming and Evolvable Machines
- Informatics and Computer Science Intelligent Systems Applications
- IS - Information Sciences
- NC - Natural Computing
- TCIAIG - IEEE Transactions on Computational Intelligence and AI in Games
- TEC - IEEE Transactions on Evolutionary Computation

Top conferences

- ACM GECCO - Genetic and Evolutionary Computation Conference
- IEEE CEC/WCCI - World Congress on Computational Intelligence
- PPSN - Parallel Problem Solving From Nature

Notes:

The tutors regularly present tutorials on Diversity Preservation at top conferences in the field, such as GECCO, PPSN, and CEC. Additionally, they are involved in the organization of a workshops focused on graph-based representation for EA. Moreover, the research group is in contact with industries that actively consider exploiting evolutionary machine-learning for enhancing their biological models, for instance, KRD (Czech Republic), Teregroup (Italy), and BioVal Process (France).

The research group has also a long record of successful applications of evolutionary algorithms in several different domains. For instance, the on-going collaboration with STMicroelectronics on test and validation of programmable devices, does exploit evolutionary algorithms and would benefit from the research.

Advanced ICT solutions and AI-driven methodologies for Cultural Heritage resilience

Recent crises and disasters have affected the European citizens' lives, livelihoods, and environment in unforeseen and unprecedented ways. They have transformed our very understanding of them by reshaping hitherto unchallenged notions of the ?local? and the ?global? and putting into question well-rehearsed conceptual distinctions of ?natural? and ?man-made? disasters. Modern and high-performance ICT solutions need to be deployed in order to prevent and mitigate the effects of disasters and climate change events by enabling critical thinking and framing a holistic approach for better understanding of catastrophic events.

The objective of this Ph.D. proposal consists of the design and development of ICT-driven solutions to develop proactive strategies for risk assessment, monitoring, and preservation of Cultural Heritage. The candidate will adopt a comprehensive interdisciplinary approach, seamlessly integrating modern techniques rooted in IoT, Machine/Deep Learning, and Big Data paradigms within the realm of cultural heritage resilience. This approach transcends purely technical facets, encompassing social and cultural dimensions to provide a holistic understanding and effective solutions.

During the three years of the Ph.D., the research activity will be divided into five phases:- Survey existing literature on modern Ai-driven ICT solutions and applications in software engineering and analyze methodologies and challenges Cultural Heritage resilience.- Design and develop a data-driven digital ecosystem - i.e., distributed IoT platform - for the collection and harmonization of heterogeneous data from the real world to enable on-top advanced visualization and analysis services (e.g., Digital Twins). A multidisciplinary approach ranging from IoT paradigms to the application of Machine/Deep Learning methodologies for Big Data analysis is required in order to allow the development of proactive strategies for risk assessment, monitoring, and preservation of Cultural Heritage.- Develop algorithms and strategies for a context-aware Cultural Heritage resilience by implementing prototype systems for evaluation and refinement.- Design and implement continuous improvement and fine-tuning strategies for the development of increasingly effective and high-performing prevention strategies.- Evaluate the effectiveness of the data-driven digital ecosystem and developed strategies through user studies and real-world projects.

Possible international scientific journals and conference: - IEEE Transactions on Computational Social Systems- IEEE Transactions on Industrial Informatics- Journal on Computing and Cultural Heritage- Journal of Cultural Heritage- Engineering Applications of Artificial Intelligence,- Expert Systems with Applications,- IEEE CoStProgramming and Object-Oriented Programming (preferable in Python).- Knowledge of web application programming.- Knowledge of IoT paradigms.- Knowledge of Machine Learning and Deep Learning.- Knowledgeof frameworks to develop models based on Machine Learning and Deep Learning Model-  internat. Conf.- IEEE SKIMA internat. Conf.

Monitoring systems and techniques for precision agriculture

Designing heterogeneous digital/neuromorphic fog computing systems and development framework 

Cloud at the edge: creating a seamless computing platform with opportunistic datacenters

https://netgroup.polito.it 

Project website: https://liqo.io

AI-driven cybersecurity assessment for automotive

Applications of Large Language Models in time-evolving scenarios

Context
Large Language Models (LLMs) have emerged as disruptive Artificial Intelligence technologies supporting a variety of Natural Language Generation tasks among which question answering, text summarization, and text paraphrasing [1,2]. Recently proposed LLMs such as LLaVA [3] support visual content as part of the LLM prompts beyond the raw text. LLMs are known to potentially suffer from the bias due to the inherent properties of the training examples. To overcome these "harms", various strategies such as in-context learning, probing, and fine-tuning have been proposed.

Research objectives
The PhD proposal has the twofold aim to address the limitations of LLMs in coping with time-evolving scenarios and timestamped data:
1)    Apply LLMs in time-evolving scenarios: Several textual and visual data sources are, by design, time-evolving. Capturing their temporal evolution is relevant to address several tasks such as intent recognition [4] and summarization [5]. The research activities will investigate the design and development of innovative LLM-based approaches to solve time-evolving tasks. 
2)    Dealing with timestamped data:  Classical LLMs are designed to handle textual data. Recent Multimodal LLMs handle visual content as well. Conversely, only a limited body of work has focused on coping with timestamped data such as time series [6]. The research activities will study new LLM-based solutions to handle timestamped data. 

Tentative work plan
1) Application of LLMs in time-evolving scenarios:
- Analysis of the state-of-the-art of LLMs and Multimodal LLMs;
- Identification of a selection of time-evolving NLP and Multimodal Learning tasks and related benchmarks. Exploration of state-of-the-art models' performance;
- Proposal of new LLM-based approaches to solve the selected tasks.
2) LLMs and timestamped data:
- Review of existing LLM-based approaches to time series and temporal sequences. Classification of their strengths and weaknesses;
- Identification of a selection of tasks related to time series data (e.g., forecasting, segmentation, classification, anomaly detection);
- Design and development of innovative LLM-based approaches to solve the selected tasks.
 
Industrial collaborations
This research activities will be partly carried out in collaboration with Amazon Research Center in Turin.
 
List of possible publication venues
- Conferences: ACM Multimedia, KDD, ACL, COLING, IEEE ICDM, ECML PKDD, ACM CIKM, INTERSPEECH, IEEE ICASSP
- Journals: IEEE TKDE, ACM TKDD, IEEE TAI, ACM TIST, IEEE/ACM TASLP
 
References
[1] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.
[2] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, AdinaWilliams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aur?lien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288, 2023.
[3] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, Jianwei Yang: LLaVA-Grounding: Grounded Visual Chat with Large Multimodal Models. CoRR abs/2312.02949 (2023)
[4] Patcharapruek Watanangura, Sukit Vanichrudee, On Minteer, Theeranat Sringamdee, Nattapong Thanngam, Thitirat Siriborvornratanakul: A Comparative Survey of Text Summarization Techniques. SN Comput. Sci. 5(1): 47 (2024)
[5] Henry Weld, Xiaoqi Huang, Siqu Long, Josiah Poon, Soyeon Caren Han: A Survey of Joint Intent Detection and Slot Filling Models in Natural Language Understanding. ACM Comput. Surv. 55(8): 156:1-156:38 (2023)
[6] Azul Garza, Max Mergenthaler Canseco: TimeGPT-1. CoRR abs/2310.03589 (2023)

Building Adaptive Embodied Agents in XR to Enhance Educational Activities

In the evolving landscape of AI and educational technology, the integration of MANNs and gamification in ECAs offers new opportunities to push the field of AI agents and create highly interactive, adaptive and engaging learning experiences in XR. The use of MANNs allows ECAs to store and recall previous interactions, resolving different problems of actual conversational agents and enabling unprecedented levels of personalized content and engagement. Ultimately, these ECAs can provide an enhanced learning experience that is both dynamic and responsive to learners' individual needs. In collaborative learning scenarios, these ECAs should be designed to act not just as facilitators but as active participants, encouraging group interaction and the overall learning process. The integration of these technologies offers the potential to explore new educational methodologies that align with the evolving digital competencies of today's learners.
 
RESEARCH OBJECTIVES:

1. Enhancing ECAs with MANNs:
- Develop ECAs that integrate with MANNs to provide a personalized learning experience by remembering and leveraging the learner's individual interactions and history.
- Explore how these memory functions can be optimized to adapt to different learning styles and preferences.
- Explore the possibilities of integrating the ECAs digital memory with emotional models to deliver lifelike interactions between agents and between the user and the agents
- In the specific context of XR-based learning, address the main challenges of MANNs, such as the limited storage capacity of certain types of networks, the complexity associated with managing external memory structures and their computational overhead, the problems associated with memory recall processes, and the efficient use of memory to store and retrieve information over extended periods of time.

2. Facilitate collaborative learning through ECAs:
- Develop ECAs that can dynamically participate in collaborative learning environments and contribute to and facilitate group-based educational activities.
- Investigate the effectiveness of ECAs in promoting group dynamics and enhancing the collaborative learning experience.
 
3. Challenges in development and implementation:
- Address Overcome the technological challenges associated with developing complex ECAs that can integrate advanced MANNs and gamification features.
- Ensure the accessibility and effectiveness of these ECAs in a wide range of educational environments, including those with limited technological resources.
- Propose novel methodologies and techniques for the MANN creation, for example or example exploit XR environments, generative AI and gamification concept to support data collection and/or models training and instruction. 

4. Assessment:
- Evaluate the impact of MANN-enhanced ECAs on the overall learning experience in XR environments.

 
The multidisciplinary nature of this project includes expertise from the fields of AI, neuroscience, psychology, education and game design. A key advantage lies in the existing partnership of the proposer's research group with the Department of Neuroscience of the Faculty of Psychology of the University of Turin, which can provide invaluable insights and contributions, especially in the field of cognitive processes and neural mechanisms, enriching the depth and applicability of the project in the field of educational technologies.
 
WORKPLAN
Year 1: Foundation and State-of-the-Art Review
- Q1-Q2: Conduct a comprehensive literature review on the current state of AI, Memory-Augmented Neural Networks (MANNs), Embodied Conversational Agents (ECAs), neuroscience, learning and education. Identify gaps in the current research and develop a detailed research proposal addressing these gaps.
- Q3 ? Q4: Begin preliminary development of the ECA framework, focusing on basic integration of neural memory.
Year 2: Development and Initial Testing
   Q1-Q2: Develop advanced features for the ECA, incorporating MANNs.
   Q3-Q4: Testing of the framework with a user panel, and its refinement according to users feedback.   
Year 3: Implementation, Evaluation, and Thesis Writing
   Implement the ECA in real-world individual and collaborative educational setting. Collect user data on its effectiveness, user engagement, and learning outcomes.
 
PUBLICATION VENUES
Journals: IEEE Trans. on Neural Networks and Learning Systems, IEEE Trans. on Learning Technologies, IEEE Trans. on Visualization and Computer Graphics, Neurocomputing, International Journal of Neural Systems
Conferences:    IJCAI, NeurIPS, ICONIP, AAAI, ICML, ICRA, and other conferences about the project topics             

COLLABORATIONS
The proposer's research group is collaborating with the Neuroscience Department of the Faculty of Psychology of the University of Turin, which will be involved in the project as domain expert and will help in providing insights about neural memory models, develop use cases, design the approach and support the assessment phase.

Real-Time Generative AI for Enhanced Extended Reality

Transferable and efficient robot learning across tasks, environments, and embodiments

Classical learning methods for robotic perception and control tend to target specific skills and embodiments, due to the difficulties in extracting transferable and actionable representations which are invariant to physical properties of the environment and of the robot. However, the performance of such specialized agents can be limited by low model capacity and training on relatively few examples. This can be particularly problematic when tackling complex and long-horizon tasks for which the cost of large-scale data collection on a single robot can be prohibitively high and the complexity of the policy to be learned might benefit from a more expressive function class (i.e., with a larger number of parameters).

Conversely, recent high-capacity, highly flexible machine learning models, such as vision transformers and large multimodal models, proved their worth in less constrained domains such as computer vision and NLP. In such domains, pre-training on large and diverse datasets is possible due to web-scale data availability. This results in rich ?generalist? pre-trained models enabling model fine tuning and adaptation to specific target tasks with large savings in terms of target data collection and positive transfer to new tasks and visual appearances.

A growing research line investigates the extension of high-capacity models to robotic tasks to enable complex skill learning across embodiments and modalities, thanks to the high flexibility of high-capacity architectures (e.g., GATO [1]). RobotCat [2] demonstrates how such models can be applied to solve complex robotic manipulation tasks with visually defined goals, while Open X Embodiment [3] demonstrates positive transfer for task goals specified in natural language. Octo further extends this concept by supporting multimodal goal definitions [4], while AutoRT [5] also supports multi-robot coordination. Large language models can also be employed to guide exploration and automate reward design for reinforcement learning [6].

However, these methods rely on very large numbers of parameters (i.e., in the order of billions), rendering model storage and real-time inference a challenge. This is a relevant roadblock when local execution on limited robotic hardware is required, as is often the case in open-world unstructured environments. Some of the most advanced multi-embodiment models (e.g., RT-2-X [3]) are so extensive that they cannot be stored locally and require communication with cloud environments to perform inference. Even more so when model fine-tuning or open-ended learning are required for tackling new tasks. Impractical computational and communication costs and catastrophic forgetting of previous tasks indeed represent a major challenge.

The objective of this project is the development of efficient methods for training, transfer, and inference of generalist high-capacity models for embodied and robotic tasks. Several approaches will be investigated, including the use of knowledge distillation, recent fine-tuning methods which proved to reduce the cost of execution of robotic policies (i.e., RT-2-X) from quadratic to linear while retaining performance levels [7], and approximation methods to reduce the number of parameters while retaining approximation power [8]. Moreover, constraining model output space to low-dimensional manifolds arising from the physical constraints of the target problem also holds promise to improve policy efficiency and safety [9] [10].

Potential publication venues include major AI, ML, robotics, and computer vision venues (e.g., TRO, RAL, TPAMI, JMLR, ICRA, IROS, CoRL, NeurIPS, ICML, ICLR, etc.)

Preliminary Main Activities Plan- M1-M4 Literature review on foundation models for robot learning- M3-M7 Empirical analysis of state-of-the-art methods for improving foundation model efficiency- M8-M15 Design and development of novel efficient methods focusing on robotic requirements and resource constraints- M16-M22 Experimental evaluation of the proposed methods- M23-M28 Development of novel methods incorporating output space constraints to enforce safety requirements while retaining efficiency and predictive capabilities- M28-M32 Experimental validation and dissemination of the results- M32-M36 Thesis writing

References
[1] Reed, Scott, et al. "A generalist agent." Transactions on Machine Learning Research (2022).
[2] Bousmalis, Konstantinos, et al. "RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation." Transactions on Machine Learning Research (2023).
[3] Padalkar, Abhishek, et al. "Open x-embodiment: Robotic learning datasets and rt-x models." arXiv preprint arXiv:2310.08864 (2023).
[4] Team, Octo Model, et al. "Octo: An open-source generalist robot policy." (2023).
[5] AutoRT: Embodied Foundation Models for Large Scale Orchestration of Robotic Agents https://auto-rt.github.io/static/pdf/AutoRT.pdf
[6] M. Kwon, S. M. Xie, K. Bullard, and D. Sadigh, ?Reward design with language models,? in Proc. Int. Conf. Learn. Representations, 2023, pp. 1?18.
[7] Leal, Isabel, et al. "SARA-RT: Scaling up Robotics Transformers with Self-Adaptive Robust Attention." arXiv preprint arXiv:2312.01990 (2023).
[8] Xiong, Yunyang, et al. "Nystr?mformer: A nystr?m-based algorithm for approximating self-attention." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 16. 2021.
[9] Liu, Puze, et al. "Robot reinforcement learning on the constraint manifold." Conference on Robot Learning. PMLR, 2022.
[10] Duan, Anqing, et al. "A structured prediction approach for robot imitation learning." The International Journal of Robotics Research (2023).

Neural Network reliability assessment and hardening for safety-critical embedded systems

NNs are increasingly adopted in the area of embedded systems, even for safety-critical applications (e.g., in the automotive, aerospace and robotics domains), where the probability of failures must be lower than well-defined (and extremely low) thresholds. This goal is particularly challenging, since the hw used to run the NN often corresponds to extremely advanced devices (e.g., GPUs, or dedicated AI accelerators), built with highly sophisticated (and hence less mature) semiconductor technologies. On the other side, NNs are known to own some intrinsic robustness, and can tolerate a given number of faults inside the hardware. Unfortunately, given the complexity of the NN algorithms and of the underlying architectures, an extensive analysis to understand which (and how many) faults are particularly critical is difficult to perform, at least when usual computational resources are available. The planned research activities aim first at exploring the effects of faults affecting the hardware of a GPU/AI accelerator supporting the NN execution. Experiments will study the effects of the considered faults on the results produced by the NN. This study will mainly be performed resorting to fault injection experiments. In order to keep the computational effort reasonable, different solutions will be considered, combining simulation- and emulation-based fault injection with multi-level one. The trade-off between the accuracy of the results and the required computational effort will also be evaluated. Based on the gathered results, hardening solutions acting on the hardware and/or the software will be devised, aimed at improving the resilience of the whole application with respect to faults, and thus matching the safety requirements of the target applications. 

The proposed plan of activities is organized in the following phases (for each phase, the indicative time span in months from the beginning of the PhD period is reported):
 - phase 1 (M1 to M6): the student will first study the state of the art and the literature in the area of NNs, their implementation on different platforms (including CPUs, GPUs, and hardware accelerators) and their applications. At the same time, the student will become familiar with existing fault injection environments (e.g., NVbitFI). Suitable cases of study will also be identified, whose reliability and safety could be analyzed with respect to faults affecting the underlying hardware.
 - phase 2 (M7-M18): suitable solutions to analyze the impact of faults on the considered accelerator will be devised and prototypical environments implementing them will be put in place.
 - phase 3 (M19-M24): based on the results of a set of fault injection campaigns performed to assess the reliability and safety of the selected cases of study, a detailed analysis leading to the identification of the most critical faults/components will be carried out.
 - phase 4 (M25 to M36): suitable hardening solutions will be proposed and evaluated.

Phases 2 to 4 will include dissemination activities, based on writing papers and presenting them at conferences (e.g., ETS, VTS, IOLTS, DATE). The most relevant proposed methods and results will be submitted for publication on the journals in the field, such as the IEEE Transactions on Computers, CAD, and VLSI, as well as Elsevier Microelectronics & Reliability.

We also plan for a strong cooperation with the researchers of other universities and research centers working in the area, such as the University of Trento, the University of California at Irvine (US), the Federal University of Rio Grande do Sul (Brazil), NVIDIA.

Design of an integrated system for testing headlamp optical functionalities

https://www.dauin.polito.it/it/la_ricerca/gruppi_di_ricerca/grains_graphics_and_intelligent_systems

https://www.italdesign.it/services-electric-and-electronics/harness-and-lighting/

Machine unlearning

In recent years, the rapid advancement of machine learning models, particularly Large Language Models (LLMs), has raised significant concerns regarding privacy and intellectual property rights. The need for responsible AI practices has become increasingly evident, driven by legal frameworks such as the General Data Protection Regulation (GDPR) that mandates the Right To Be Forgotten. Additionally, high-profile cases involving LLMs have highlighted the need to address issues related to the unintentional retention of sensitive information and potential copyright infringements.

The proposed research activity on Machine Unlearning (MU) aims to tackle these challenges by developing novel techniques to selectively erase or modify previously acquired knowledge from machine learning models. The primary objectives of this research are twofold: first, to explore the current state of the art in MU, and second, to propose innovative architectures, algorithms, and evaluation metrics to enhance the efficacy of the unlearning process. Through these goals, the aim is to contribute to the establishment of ethical and responsible AI practices, ensuring compliance with legal requirements and mitigating the risks associated with unintentional information retention by machine learning models.

The workplan for this PhD is structured to comprehensively address the multifaceted challenges of MU. The research will focus on proposing novel architectures and algorithms that facilitate effective unlearning while preserving the model's overall performance. Given the current lack of definitive metrics for MU, part of the research efforts will be focused toward trying to identify more suitable and comprehensive metrics.

The research activity progresses from foundational research to the practical implementation, validation and application of MU techniques. An outline of the possible research plan is as follows.

- First year
The first year will be dedicated to literature review and conceptualization, leading to the formulation of the main research objectives for the rest of the doctorate. This initial phase involves an extensive study of the literature, identifying gaps and shortcomings ? leading to the definition of initial proposals for improvements over state-of-the-art techniques.

- Second year
Based on the areas of opportunity identified and the preliminary proposals made, the candidate will work on the ideation and implementation of novel architectures and algorithms for MU, with ongoing validation and refinement based on the feedback obtained from experiments and evaluations.

-Third year
The final year will focus on consolidating the findings and defining the applications of main interest for the output produced.
During the second/third year, the candidate will have the opportunity to spend a period of time abroad in a leading research center.
Publication venues for this research include leading conferences and journals in the fields of machine learning and artificial intelligence. Key conferences include the conference on Neural Information Processing Systems (NeurIPS), the International Conference on Machine Learning (ICML), and the International Conference on Representation Learning (ICLR). Additionally, reputable journals such as the Journal of Machine Learning Research (JMLR) and the IEEE Transactions on Neural Networks and Learning Systems will be sought for in-depth dissemination of research contributions.

Test, reliability, and safety of intelligent and dependable devices supporting sustainable mobility

Research objectives
The novelty of this research lies in its focus on sustainable mobility, which is an emerging area of research with great potential for real-world impact. The work is expected to significantly improve the reliability and safety of electronic systems, thereby enhancing the performance of safety-critical applications. The research team's expertise in electronic design automation (EDA) will be leveraged to develop robust methodologies that are both practical and effective. Furthermore, this research is aligned with the goals of the National Centers on Sustainable Mobility and HPC, as well as the Extended Partnership on Artificial Intelligence, which further emphasizes its significance in advancing the state-of-the-art in this field.

The objectives of this research are summarized as follows:- Identify a suitable hardware platform for sustainable mobility applications with particular emphasis on RISC-V based systems.- Identify suitable software for mobility applications to be used as a representative benchmark for the qualification activities.- Assess dependability figures on the identified hardware/software infrastructure to identify critical parts of the design that require hardening.- Develop innovative hardening solutions to improve the reliability of critical areas in the design.- Focus on sustainable mobility as an emerging area of research with great potential for real-world impact.- Establish a comprehensive assessment methodology in collaboration with EDA partners.

Outline of possible research plan

First year: 
The candidate will start by conducting a thorough literature review on dependable electronic systems and sustainable mobility to identify the most recent and relevant research works. They will then select a suitable hardware platform for sustainable mobility applications, taking into account a variety of RISC-V based systems publicly available, and using IP cores from industrial partners (e.g., Synopsys). Furthermore, they will identify suitable software for mobility applications, including AI applications, and leverage publicly available benchmarks developed for other domains such as automotive and space. The candidate will develop a preliminary assessment methodology for the identified hardware and software infrastructure, which will be refined and improved in the following years.

Second year: 
The candidate will focus on identifying the critical parts of the design that require hardening and implementing initial solutions to enhance the overall system reliability. They will perform dependability analysis on the identified hardware/software infrastructure, aiming to improve the quality of the developed assessment framework. The candidate will explore various hardening techniques, including redundancy, error-correcting codes, and fault-tolerant architectures, and select the most suitable ones to enhance the system reliability and safety.

Third year: 
The candidate will develop innovative hardening solutions to improve the reliability of critical areas in the design, while ensuring the availability of safety mechanisms in the event of a fault. They will develop a comprehensive assessment methodology in collaboration with EDA partners, leveraging their expertise in electronic design automation to refine and optimize the assessment process. The proposed methodologies will be extensively evaluated through simulations and testing, and the candidate will collaborate with industry partners to validate their effectiveness on real-world applications.

List of possible venues for publications
The candidate will prepare and submit papers to top-tier conferences and journals in the field of electronic systems, embedded systems, and fault tolerance.

Possible venues for publications could include:- IEEE Transactions on Computers - IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems- IEEE Transactions on Very Large Scale Integration (VLSI) Systems- International Conference on Computer-Aided Design (ICCAD)- International Test Conference (ITC)- IEEE European Test Symposium (ETS)- Design, Automation and Test in Europe Conference (DATE)- RISC-V Summit

Projects
The research is consistent with the themes of the National Centers on Sustainable Mobility and HPC, as well as with those of the Extended Partnership on Artificial Intelligence, in which members of the CAD group participate.
The research will be supported by industrial partners involved in active collaborations. Synopsys is involved in research activities on functional safety and reliability, and provides licensed tools, IP cores, and support. Infineon is also involved in the frame of research contracts on electronic system dependability.

Cybersecurity for a quantum world

Hard security is typically based on mathematical cryptographic algorithms that support computation of symmetric and asymmetric encryption, key exchange, digital signature, and hash values.
 
Several of these algorithms (e.g. RSA, ECDSA, ECDH) are threatened by the advent of quantum computing. NIST and other bodies have thus selected new quantum-resistant algorithms and advocated their fast adoption in current security solutions. However this is not a simple change, as there are several intertwined aspects to be considered, such as hardware support, key-length, and X.509 certificates.
 
Purpose of this work is to evolve various security components of modern ICT infrastructures to quantum-resistant versions. This may include secure network channels (e.g. TLS, IPsec), digital signatures, secure boot, Trusted Execution Environment (TEE), and remote attestation.
 
The gross objective is the design and test of quantum-resistant versions of several security solutions in an open-source environment (e.g. Linux, Keystone). 
 
The specific objectives of this research activity are:
1. Identify security components threatened by quantum computing and review proposed standards to make them quantum-resistant.
2. Extend existing open-source systems and components (e.g. Linux, Keystone, openSSL, mbedTLS could be suitable targets) to support the proposed quantum-resistant solutions.
3. Implement a system with the hardware and software components needed to demonstrate the feasibility and performance of the improved quantum-resistant elements.
 
The first year will be spent studying the existing security paradigms and how they are affected by quantum computing. The PhD student will also analyse the proposed post-quantum algorithms and evaluate their performance and hardware requirements. During this year, the student should also follow most of the mandatory courses for the PhD and submit at least one conference paper.
During the second year, the PhD student will design a custom approach for quantum-resistant secure channels and trusted execution environment, possibly enriched with specialized hardware elements. At the end of the second year, the student should have started preparing a journal publication on the topic and submit at least another conference paper.
Finally, the third year will be devoted to the Implementation and evaluation of the proposed solution, compared with the existing ones. At the end of this final year, a publication in a high-impact journal shall be achieved.
 
Possible target publications: IEEE Security and Privacy, Springer International Journal of Information Security, Elsevier Computers and Security, Future Generation Computer Systems.

This research is part of the Horizon Europe QUBIP project (Quantum-oriented Update to Browsers and Infrastructures for the PQ Transition) https://quibip.eu/

Bridging Human Expertise and Generative AI in Software Engineering

Research objectives

The main objectives of the PhD programme are the following:The identification of generative AI mechanisms that can aid in code generation from software requirements. The development and assessment of methods for the evaluation of the correctness and the dependability of the application of generative AI to code development;The conduction of formal experiments to evaluate how code generated by AI Compares to human-written code in both functional and non-functional terms.

Outline of the research work plan

Task 1: Preliminary evaluation of state-of-the-art solutions (M1-M3)

The task involves a comprehensive assessment of current solutions in the domain. The objective is to evaluate existing methodologies, technologies, and frameworks relevant to the research context. This preliminary analysis will be conducted systematically by applying Kitchenham's guidelines for conducting Systematic Literature Reviews in the Software Engineering research field. The systematic literature review will also consider grey literature sources (i.e., non-peer-reviewed sources available on various internet sources) to cope with the high novelty of the generative AI research field. The systematic evaluation of the state of the art will be complemented with open and structured interviews with practitioners and developers to understand their main needs and most common practices.

Task 2: Selection and Integration of Generative AIs for Code Generation (M4-M18)

This task focuses on selecting, customizing, integrating, and training a Generative AI, specifically a Large Language or Foundation Model, to generate code from formal requirements. It includes understanding various use case and formal requirement languages and creating modules for translating natural language requirements into structured notations like Use Case Diagrams. The process involves preprocessing data -collecting, cleaning, and structuring use case language datasets- and training the AI to understand these scenarios. Ongoing evaluation and refinement of the AI are crucial for accuracy. The main goal is to develop a solution that translates use case specifications into high-quality code, with evaluations based on the development effort, error rates, and requirement alignment. The task will use Software Repository Mining (MSR) techniques for diverse dataset collection. The implementation phase of this research will follow the Agile Software Development practices, streamlining the entire software development lifecycle to assess the efficacy of AI-generated code against existing tools and manually written code for both front-end and back-end applications. Furthermore, the research is dedicated to pioneering methods for automatically generating synchronized documentation and unit testing. It will also investigate strategies for conducting code quality reviews, monitoring resource usage efficiently, and evaluating the software's business impact, thereby tailoring the development process to meet the demands of network operations.

Task 3: Definition of assessment methods for Generative AI-based code development (M13-M24)

The task focuses on defining robust methods to assess Generative AI-based code development. This entails the definition of structured procedures to assess the accuracy, reliability, and compliance with the requirements of the generated code. The goal is to establish a rigorous framework for ensuring the quality of code produced by Generative AI, thus advancing the state of the art in Generative AI code development.
Task 3 will involve applying systematic techniques to build taxonomies in Software Engineering (ref. Paul Ralph) through the Straussian Grounded Theory technique (ref. Strauss).

Task 4: Analysis of the non-functional implications of Generative AI-based code development (M22-36)

The task focuses on a comprehensive analysis of the non-functional implications inherent to Generative AI-based code development. This includes scrutinizing factors such as scalability, performance, readability and maintainability of the generated code. The objective is to discern and mitigate any adverse effects of integrating Generative AI into the code development process.
The task will involve conducting empirical experiments over statistically significant samples to compare non-functional properties of software generated by human developers, software obtained through Generative AI, and software obtained through a synergistic interaction between human developers and generative AI tools.

List of possible venues for publications

The target for the PhD research includes a set of conferences in the general area of software engineering (ICSE, ESEM, EASE, ASE, ICSME) as well as in the specific area of testing (ICST, ISSTA).
More mature results will be published in software engineering journals, mainly IEEE Transactions on Software Engineering, ACM TOSEM, Empirical Software Engineering, Journal of Systems and Software, and Information and Software Technologies.
 

Explaining AI (XAI) models for spatio-temporal data

Machine learning models are increasingly adopted to assist human experts in decision-making. Especially in critical tasks, understanding the reasons behind machine learning model predictions is essential for trusting the model itself. For example, experts can detect model wrong behaviors and actively work on model debugging and improvement. Unfortunately, most high-performance ML models lack interpretability.
 
The research activity will consider application domains (e.g., transportation, industry, medical care, climate) in which the availability of understandable explanations is particularly relevant for explaining anomalous behaviors. The explanation algorithms will target different types of spatio-temporal data (e.g., multivariate time series, spatiotemporal graphs, trajectories, spatio-temporal matrices). The following different facets of XAI (Explainable AI) will be addressed.
 
Model understanding. The research work will address local analysis of individual predictions. These techniques will allow the inspection of the local behavior of different classifiers and the analysis of the knowledge different classifiers are exploiting for their prediction. The final aim is to support human-in-the-loop inspection of the reasons behind model predictions.
 
Model trust. Insights into how machine learning models arrive at their decision allow evaluating if the model may be trusted. Methods to evaluate the reliability of different models will be proposed. In case of negative outcomes, techniques to suggest enhancements of the model to cope with wrong behaviors and improve the trustworthiness of the model will be studied.
 
Model debugging and improvement. The evaluation of classification models generally focuses on their overall performance, which is estimated over all the available test data. An interesting research line is the exploration of differences in the model behavior, which may characterize different data subsets, thus allowing the identification of potential sources of bias in the data.

PhD years organization
YEAR I: state-of-the-art survey for algorithms and for XAI both for time series and spatio-temporal data considering, e.g., feature attribution-based explanations, attention-based explanation, and counterfactual explanation; performance analysis and preliminary proposals of improvements over state-of-the-art algorithms; exploratory analysis of novel, creative solutions for XAI; assessment of main explanation issues in 1-2 specific industrial case studies. 
 
YEAR 2: new algorithm design and development; experimental evaluation on a subset of application domains considering public domain datasets (e.g. in the transportation field, TaxiNYC, METR-LA, PEMS-BAY, in the healthcare field, PTB-XL, NYUTron, and MIMIC-III, in the financial field, StockNet and NASDAQ-100); deployment of algorithms in selected industrial contexts. 
 
YEAR 3: algorithms improvements, both in design and development, experimental evaluation in new application domains.
 
During the second-third year, the candidate will have the opportunity to spend a period of time abroad in a leading research center.
 
Publication venues for this research include leading conferences and journals in the fields of spatio-temporal data managent, machine learning and artificial intelligence: 
 
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TIST (Trans. on Intelligent Systems and Technology)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Machine Learning with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
 
IEEE/ACM International Conferences (e.g., ACM KDD, ACM SIGSPATIAL, IEEE ICDM, NeurIPS)

Advanced data modeling and innovative data analytics solutions for complex application domains

The PhD student will work on the study, design and development of proper data models and novel solutions for the integration, storage, management and analysis of big volumes of heterogeneous data collections in complex application domains. The research activity involves multidisciplinary knowledge and skills including database, machine learning and artificial intelligence algorithms, and advanced programming. 
 
Different application contexts will be considered to highlight a wide range of data modeling and analysis problems, and thus lead to the study of innovative solutions. The objectives of the research activity consist in identifying the peculiar characteristics and challenges of each considered application domain and devise novel solutions for the modelling, management and analysis of data for each domain. Example scenarios are urban context and in particular urban mobility, and the medical domain. 

 More in detail, the following challenges will be addressed during the PhD:
 
1. Modeling Heterogeneous Data: Design innovative approaches for modeling heterogeneous data, including structured and unstructured data from different sources, integrating them into a single coherent framework. The experience gained on data modeling in different application contexts can lead to the realization of a Computer-Aided Software Engineering (CASE) tool that guides the user through the design process, reducing design time and improving the quality of the modeling result.
 
2. Innovative algorithms for data analytics. Study, design, and implementation of innovative machine learning algorithms, with a primary emphasis on clustering and classification tasks. The objective is to overcome limitations of current approaches, enhancing their accuracy, scalability, and ability to deal with heterogeneous data collections. 
 
3. Scalable Learning: Investigate scalable learning techniques to address the increasing complexity and volume of data for achieving optimal performance in big data environments. This research is indeed driven by the growing demand to develop machine learning systems capable of dynamically adapting to the increasing complexity of data and models. For recent machine learning/AI applications, it is crucial to propose innovative models capable of handling large volumes of data with parallel and scalable solutions.
 
The research activity will be organized as follows.
1st Year. The PhD student will start considering a first reference application domain (for example the urban scenario) and a first reference use case in this scenario (for example urban mobility). The PhD student will review the recent literature in the selected use case to (i) identify the most relevant open research issues, (ii) identify the most relevant data analysis perspectives for gaining useful insights, and (iii) assess of main data analysis issues. The PhD student will perform an exploratory evaluation of state-of-the-art technologies and methods on the considered domain, and she/he will present a preliminary proposal for the optimization techniques of these approaches.
 
2nd and 3rd Year. Based on the results of the 1st year activity, the PhD student will design and develop a suitable framework including innovative data analytics solutions to efficiently model data in the considered use case and extract useful knowledge, aimed at overcoming weaknesses of state-of-the-art methods.
 
Moreover, during the 2nd and 3rd year, the student will progressively consider a larger spectrum of application domains. The student will evaluate if and how his/her proposed solutions can be applied to the new considered domains as well as he/she will propose novel analytics solutions.
 
During the PhD, the student will have the opportunity to cooperate in the development of solutions applied to the research project on smart cities (e.g., PRIN project on the development of an atlas for historic buildings in an urban context). The student will also complete his/her background by attending relevant courses. The student will participate to conferences presenting the results of his/her research activity.
 
Possible pubblication venues includes international journals such as IEEE Transactions on Intelligent Transportation Systems, Information Systems Frontiers (Springer), Information sciences (Elsevier), and international conferences such as IEEE Big data, 
ACM Inter. Conf. on Information & Knowledge Management (CIKM), IEEE International Conference on Data Mining (ICDM)

Functional Safety Techniques for Automotive oriented Systems-on-Chip

The phd student will pursue objectives in the broader research field of the Automotive Reliability and Testing.
 
Key enabling factor for this work is the availability of a setup that includes both netlists to be simulated and real silicon chips with development boards to use effectively. 
 
Techniques for developing Software-Based Self-Test libraries 
In this research field, the PhD student will look in the following directions:
o  Creation of benchmarks design to use along the studies.
    1.     RISC5-oriented SoC
    2.     Industrial benchmarks provided by industrial supporters, including netlist and silicon implementation
o  Identification of the current industrial solutions for the development of SBST libraries and improvement of the state-of-the-art
    1.     Usage of currently available tools designed by EDA and classification of strength and weaknesses of commercial tools
    2.     Creation of "wrappers" to evolve EDA ecosystems for custom data collection and improved generation flows
    3.     Investigation on the extension of existing tools by creating ad-hoc tools adding analytical value and accelerating SBST development.
o  Data collection tools to enrich the log capabilities of EDA tools
o  High-level techniques based on silicon

Techniques for grading and developing System-level Test (SLT) libraries
Research in the field of System-Level Test by addressing the open points of this modern testing technique. SLT is a strong demand and quite an obscure subject that mainly concerns the integration of heterogeneous systems. Development of methods for achieving high SLT coverage will include:
o  Coverage computation techniques including simulation and silicon-based methods 
o  Generation of SLT methods that maximize the coverage levels.

The working plan for the PhD student is recalling the objectives drawn in the previous sections. The order is not fixed and may vary according to the advancement during the PhD program.
 
1st year
1.     Benchmarks design including simulation and test synthesis 
2.     Preliminary usage of EDA tools to measure SBST and STL coverage
3.     Identification of weaknesses and plan a list of counteractions
2nd year
1.     Ad-hoc tools preliminary development upon results collected during the 1st year
2.     Codification of methods to increase the coverage and speed up the generation process
3rd year
1.    Exploration of alternative coverage measurements for SBST and SLT 
Completion of a flexible and extensible environment to quick creation of SBST and SLT libraries.

Human-aware robot behaviour learning for HRI

This proposed PhD thesis aims to explore the domain of learning human skills from egocentric videos and transferring them to robotic systems. The increasing integration of robots into various aspects of human life highlight the need of developing a more intuitive and adaptive approach to skill acquisition, able to learn complex skills and adapt to various scenarios, where traditional learning paradigms fail. Egocentric videos, captured from a first-person perspective, provide a rich and unique source of contextual information that can enhance the learning process. This research seeks to leverage this unique sensing approach to develop a framework for transferring acquired human skills to robots, enabling them to perform complex tasks in diverse environments. 

Major Objectives:
-      Human Behavior understanding from ego- and exo-vision: Investigate methods for extracting relevant information from egocentric videos, eventually in combination with third person perspective, focusing on understanding human actions, interactions, and environmental cues.
-       Skill Representation: Develop a robust representation of human skills extracted from egocentric videos, considering both spatial and temporal aspects to capture the dynamics of actions.
-       Transfer Learning to Robots: Design a transfer learning framework that enables the adaptation of learned human skills to robotic systems, accounting for differences in morphology, sensors, and actuators.
-       Adaptive/Continual Learning: Explore techniques for adaptive and continual learning, allowing robots to continuously refine their skills through interaction with the environment and human feedback.
-       Real-world Applications: Evaluate the proposed framework in real-world scenarios, such as assistive robotics, industrial automation, and healthcare, to assess its practicality and generalizability.

Methodology:
The research methodology involves a combination of computer vision, machine learning, and robotics techniques. Deep learning models will be employed for ego-vision learning and skill representation, while transfer learning techniques will be investigated to adapt these representations to robotic platforms. Real-world experiments will validate the effectiveness and efficiency of the proposed framework.

Significance:
This research addresses a critical gap in the field of robotics by focusing on intuitive skill transfer from humans to robots, enhancing their adaptability and autonomy. The outcomes of this thesis will contribute to the development of more versatile and capable robotic systems, fostering their integration into various domains.

Keywords: Egocentric Videos, Skill Transfer, Robotics, Transfer Learning, Human-Robot Interaction.

Generative and Adversarial AI for Testing and Optimization of Connected and Automated Vehicles

Digital twin (DT) and Mixed Reality (MR) systems and technologies can assist Connected and Automated Vehicles (CAVs) in using advanced sensors, like ultrasonic radars, cameras, and LiDAR, to gather data from their surroundings and create virtual representations. Artificial intelligence (AI) methods then use these virtual models to inform driving decisions, enhancing CAVs' responses to dynamic road conditions. However, CAVs still face limitations in environmental perception due to occlusions and other sensor limitations, even with high-class LiDAR and panoramic cameras. Leveraging inter-vehicle connectivity and Road Side Units (RSU) communicating with vehicles can alleviate these problems.
To overcome these challenges, there is a need for a system where CAVs, roadside units (RSUs), and virtual simulators collaboratively share and fuse sensing data to achieve comprehensive environmental perception. Realizing this large-scale data collection and processing for real-time driving simulation and training of CAVs usually means developing virtual traffic and driving simulation platforms. These platforms leverage DT and MR technologies to create virtual representations of AVs, facilitating efficient traffic and training data collection and allowing for simulation and testing of rare scenarios, like virtual traffic accidents and car collisions, under realistic conditions.
However, traditional testing methods may not cover all possible scenarios, especially unusual, or extreme, conditions. Generative AI offers a solution by creating synthetic data and scenarios, while adversarial AI can help identify weaknesses. Combining these approaches has the potential to provide a powerful framework for comprehensive testing and optimization. 
Research Objectives

 (1) Development of Generative Models:Design and train generative models (e.g., GANs, VAEs) to create diverse driving scenarios.Generate synthetic datasets (or integration of existing datasets using synthetic data) including varying weather conditions, traffic/movement patterns, and unexpected obstacles.
(2) Adversarial AI Implementation:Develop methods to create adversarial examples that challenge the vehicle's perception and decision-making processes, identifying the types of scenarios that are most likely to cause failures or suboptimal performance in CAVs.Investigate the use of Adversarial AI for the setting and optimization of system parameters.
(3) Integration and Adaptive Testing:Combine generative and adversarial methods to create a dynamic testing environment, based on simulation or hardware-in-the-loop emulationImplement adaptive testing where generative models produce scenarios based on adversarial findings.Continuously refine and improve CAV systems based on testing outcomes.
(4) Evaluation:Conduct extensive simulations to evaluate the effectiveness of the proposed framework.Analyze the impact on safety, reliability, and performance of CAVs.Compare results with traditional testing methods to highlight improvements.
Research Work Plan
- Year 1: Literature review, development of generative models, evaluation of suitable adversarial AI techniques, initial scenario generation.- Year 2: Implementation of adversarial AI techniques, integration with generative models.- Year 3: Adaptive testing, extensive simulations and evaluation, system refinement.

Publication Venues

Conferences: IEEE Infocom, IEEE VNC, IEEE VTC, IEEE ICC/Globecom
Journals: IEEE Transactions on Vehicular Technology, IEEE Transactions on Mobile Computing, IEEE Transactions on Automation Science and Engineering

The ideal candidate should:  
- possess a strong foundation in machine learning concepts, algorithms, and frameworks 
- be proficient in programming languages such as Python and C++
- have strong development skills, including code versioning (e.g., Git), debugging, and optimization
- have experience in using ML libraries and tools such as TensorFlow, PyTorch, and Keras

Past experience in the application of AI/ML techniques in the analysis of wireless and/or mobile networks is a plus.

Continuous Machine Learning for Cybersecurity

Research objectives:
 
Investigate and evaluate the capabilities of continuous learning strategies in automating the identification and investigation of security threats. 

The candidate will perform research to determine whether, and to what extent, the recent advances in machine learning could be used to automate and assist security analysts in the process (i) of identifying new attacks and (ii) setting up countermeasures.
 
In the first phase, the candidate will study the state of the art of continuous machine learning, including the problem of data collection in distributed scenarios and federated learning solutions.

In the second phase, the candidate will investigate how and if machine learning models can extract insights from raw data to identify new threats. For this, we will first build a generic data representation using embeddings, which will be specialised to solve custom downstream tasks.  We will consider a distributed scenario, where different data sources collect data. Via federated learning, the candidate will investigate how to train a shared embedding model suitable to then solve specific tasks such as threat classification, anomaly detection, or identifying new attack patterns.
 
The candidate will consider multi-modal embeddings to conceptually constrain the embeddings towards the right task. 
By forcing the model to create multi-modal embeddings conceptually constrained to the right task, the model will possess the ability to generalize and autonomously reason about novel and previously unencountered tasks. For instance, test joint learning of (i) data collected from honeypots and (ii) the packet payload, using, e.g., contrastive learning techniques. 
 
The project will involve a collaboration with Huawei Technologies France and Politecnico di Torino. 

Outline of the research work plan

1st  year
- Study of the state-of-the-art of security log analysis and state-of-the-art continuous learning models in ML.
-  Data collection and analysis of raw and structured data on security devices such as honeypots, darknets, IDS, etc. 

2nd  year
- Adaptation and extension solutions to build a distributed platform for data collection and model training using federated learning.
- Propose and develop innovative solutions to the problems of cyber threats analysis with machine learning solutions.
- Propose multi-modal embeddings for network raw data and security logs.

3rd  year 
- Tune the developed techniques and highlight possible strategies to counteract the various threats.
- Application of the strategies to new data for validation and testing.

References:

-       Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Drago, Idilio, Milan, Giulia, Houidi, Zied Ben, Rossi, Dario (2023). Cross-network Embeddings Transfer for Traffic Analysis. IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, p. 1-13, ISSN: 1932-4537, doi: 10.1109/TNSM.2023.3329442
-       Luca Gioacchini, Luca Vassio, Marco Mellia, Idilio Drago, Zied Ben Houidi, Dario Rossi (2023). i-DarkVec: Incremental Embeddings for Darknet Traffic Analysis. ACM TRANSACTIONS ON INTERNET TECHNOLOGY, vol. 23, p. 1-28, ISSN: 1533-5399, doi: 10.1145/3595378
-       Huang, Kai, Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Dynamic Cluster Analysis to Detect and Track Novelty in Network Telescopes, 9th International Workshop on Traffic Measurements for Cybersecurity (WTMC 2024), Vienna, Austria, July 2024
-       Huang, Kai, Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Incremental Federated Host Embeddings for Network Telescopes Traffic Analysis, IEEE International Workshop on Generative, Incremental, Adversarial, Explainable AI/ML in Distributed Computing Systems (AI-DCS), Jersey City, New Jersey, USA, July 2024
List of possible venues for publications

Security venues: IEEE Symposium on Security and Privacy, IEEE Transactions on Information Forensics and Security, ACM Symposium on Computer and Communications Security (CCS), USENIX Security Symposium, IEEE Security & Privacy;

AI venues: Neural Information Processing Systems (NeurIPS), International Conference on Learning Representations (ICLR), International Conference on Machine Learning (ICML), AAAI Conference on Artificial Intelligence, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD);

Computer networks venues: Distributed System Security Symposium (NDSS), Privacy Enhancing Technologies Symposium, The Web Conference (formerly International World Wide Web Conference WWW), ACM International Conference on Emerging Networking EXperiments and Technologies (CoNEXT), USENIX Symposium on Networked Systems Design and Implementation (NSDI);

REM sleep disorder detection using tachogram and other bio-signals measured using wearable devices

Based on encouraging preliminary results related to the diagnosis of sleep disorders using lightweight, low-cost instrumentation, the research objective of this proposal is to develop ML/AI algorithms to perform a healthy/RBD classification based solely on a tachogram (sequence of time intervals between successive R peaks of the ECG signal) recorded during sleep using a wearable device (e.g., chest heart rate monitor or optical wrist reader). In fact, it is known that the heart rate variability is strictly correlated to such sleep disorder. The results can possibly be improved using multimodal data from wearable EMG and/or EEG data, to match the sensitivity requirements of screening applications (low false negative rates). A measurement champaign performed both in hospital and at home will be implemented to validate the technique. The work is supported by the neurology department and the center for sleep disorders of the Molinette Hospital, in the context of a long-lasting cooperation.
WORKPLAN: YEAR 1.
Task 0. Revision of the state of the art related to the use of tachogram for sleep disorder identification.
Task 1. Implementation of ML/DL algorithms for RBD/healthy control classification based on tachogram. The algorithms will be trained and tested using publicly available ECG data, processed following the AASM guidelines for a correct performance of the polysomnographic examination. 
Task 2: Definition of a protocol for bio-signal recording suitable for domestic use, using commercial wearable devices and in cooperation with the clinical staff the Molinette Hospital, Center for Sleep Disorders and Neurology Department. Selection of proper devices, with CE certification. Submission to the Ethics Committee of the Molinette Hospital. 
Task 3. Definition of the inclusion/exclusion criteria for the selection of (approximately) 20 patients to be included in the trial. Due to the several possible confounding factors, also patients affected by other sleep disorders (in particular, obstructive sleep apnea syndrome - OSAS) will be included in the trial to evaluate the algorithm capability of coping with confounding factors.
ACTIVITIES: YEAR 2.
Task 4. Algorithm testing and performance comparison with polysomnographic data and clinical reports provided by the Molinette Hospital. Refinement of the algorithms implemented in Task 1. 
Task 5. Study of the possible performance enhancement provided by the joint use of EMG or one EEG electrode for sleep stage definition.
Task 6. Preliminary enrollment of patients enrolled at the Molinette Hospital and measurements made in hospital, using the selected wearable devices, during the execution of the normal polysomnography. First comparison of automatic classification with the clinical reporting.
ACTIVITIES: YEAR 3.
Task 7. Finalization of the data collection champaign in hospital, and related data analysis.
Task 8. Finalization of algorithms using multi-dimensional data and their validation on newly collected data.
Task 9. Data collection at home on a limited number of patients and reporting of the possible problems related to the self-monitoring process. Comparison if the quality of data measured at home and in hospital.
Task 10. Critical analysis of the results, proposal of integration of the algorithms with heterogeneous data present in the patient's clinical records.
 
N.B: Due to the complexity and novelty of this topic, the Task description is forcedly preliminary and may be subject to modifications depending on the obtained results. 
 
We plan to have at least two journal papers published per year.
Target journals:
IEEE Transactions on Biomedical Engineering
IEEE Journal on Biomedical and Health Informatics
IEEE Access
IEEE Journal of Translational Engineering in Health and Medicine
MDPI Sensors
Frontiers in Neurology
 
COOPERATIONS:
- CNR-IEIIT
- Department of Neuroscience "Rita Levi Montalcini 
- Molinette Hospital, Neurology Department and center for Sleep Disorders

ON GOING PROJECTS ON RELATED TOPICS: 
- PNRR Ecosistema NODES Nord-Ovest Digitale e Sostenibile
- PNRR Salute Complementare ?Health Digital Driven Diagnostics, prognostics and therapeutics for sustainable Health care?
- PRIN-PNRR 2022 ?Objective monitoring of axial symptoms in Parkinson's disease: quantitative assessment in daily life based on the use of wearables, video sensing and artificial intelligence (OMNIA-PARK)?

Innovative Techniques for Synthetic Renders Automatic Comparison

Innovative Techniques for Storyboarding in Extended Reality

Innovative Robot Learning Techniques for Human-Robot Collaboration

The proposed research aims to advance the field of human-robot collaboration by developing and demonstrating innovative robot learning techniques that enable effective collaboration in dynamic environments. 
 
A collaborative robotic arm manipulator is a type of robot that is designed to work alongside humans in a shared workspace. These robots are typically used in manufacturing, where they are capable of performing tasks such as picking, placing, and manipulating objects. Collaborative robotic arm manipulators are equipped with sensors and safety features that allow them to operate safely in close proximity to humans and are designed to be easy to use and adaptable to a wide range of tasks. They are used in a wide range of fields, including manufacturing, healthcare, retail, agriculture, construction, and service industries. Overall, the use of collaborative robotic arms is growing in a wide range of fields, as they are capable of performing tasks quickly, accurately, and safely, and can improve efficiency and reduce labor costs. There are several learning techniques that are commonly used for programming and operating collaborative robotic arm manipulators: lead-through programming, point-and-click programming, scripting, and hardware-in-the-loop (HIL) simulations. 
 
A novel emerging technique is learning by imitation, which refers to the process of a robot learning to perform a task by observing and copying the actions of a human or another robot. There are several approaches to robot learning by imitation, including: - Behavioural cloning: in this approach, the robot is trained to mimic the actions of a human or another robot by learning from a dataset of demonstrations. - Learning from demonstrations: In this approach, the robot is provided with a set of demonstrations and learns a model of the task by inferring the underlying rules or principles from the demonstrations. - Human-robot interaction: In this approach, the robot learns by interacting with a human in a collaborative manner, as the human provides guidance and feedback to the robot as it performs the task.

Overall, robot learning by imitation is a powerful tool for enabling robots to learn new tasks and adapt to changing environments. The goal is to research and develop novel learning by imitation techniques and to contribute to the development of more effective and efficient human-robot collaboration systems. 
  
Active industrial collaboration pertaining to human-robot collaboration systems and applications is ongoing with Comau.

 Research objectives:- To review and analyze current state-of-the-art robot learning techniques for human-robot collaboration. This will involve a comprehensive literature review of existing approaches, including supervised and unsupervised learning, reinforcement learning, and transfer learning.- To identify the challenges and limitations of these techniques in dynamic environments. This will involve a detailed analysis of the strengths and weaknesses of existing approaches, as well as an exploration of the unique challenges posed by dynamic environments such as manufacturing and healthcare settings.- To propose and evaluate novel robot learning approaches that address these challenges and enhance human-robot collaboration. This will involve the development of new robot learning algorithms and approaches that are specifically designed to enable effective human-robot collaboration in dynamic environments.- To demonstrate the practicality and effectiveness of the proposed techniques through simulations and experiments. This will involve the implementation and testing of the proposed techniques in realistic simulation environments, as well as in real-world scenarios to evaluate their performance and effectiveness.
 
 
Research work plan:
1st year- Conduct a comprehensive literature review of existing learning by imitation techniques for human-robot collaboration with robotic arm manipulators.- Identify and analyze the challenges and limitations of these techniques in dynamic environments.- Improve programming, problem-solving, and data analysis skills pertaining to robot learning techniques and human-robot interaction - Develop strong communication skills to communicate research findings and ideas effectively to a range of audiences
2nd year- Develop and propose novel learning approaches (based on the learning by imitation technique) that address existing challenges and enhance human-robot collaboration.
3rd year- Implement and test the proposed techniques in simulation and real-world scenarios.- Analyze and evaluate the results of the simulations and experiments.
 
Possible venues for publications:- IEEE Transactions on Robotics (Q1, IF 6.835)- International Journal of Human-Computer Interaction (Q1, IF 4.920) - ACM Transactions on Human-Robot Interaction (Q2, IF 4.69)- IEEE International Conference on Human-Robot Interaction (HRI)- International Conference on Robotics and Automation (ICRA)

Adaptive, Agile and Automated Cybersecurity Management

Security of Linux Kernel Extensions for Fast Packet Processing

Smart digital technologies for advanced and personalised health monitoring

In an increasingly ageing society, the prevalence of health-related problems (e.g. cardiovascular complications) is rapidly increasing, as is the incidence of chronic and neurodegenerative disorders. In recent decades, the development of new digital technologies promises to revolutionise the management of diseases and health-related issues. In particular, digital solutions are being designed and optimised to facilitate early diagnosis, unconstrained continuous monitoring and evaluation of disease progression and treatment effectiveness. These new technologies include wearable sensors that can be worn on the body for long periods of time and enable minimally invasive, unconstrained, continuous and accurate real-time health monitoring. In addition, devices such as RGB cameras, infrared sensors and radar systems offer the advantage of non-contact monitoring of vital signs and human movement. Non-invasive data collection procedures ensure the ecological validity of the recorded data, promoting patient comfort and compliance and thus enabling long-term monitoring. At the same time, the huge amount of multi-modal data generated by different sensors can pose challenges for data analysis, fusion, processing, synthesis, interpretation, security and privacy. To overcome these challenges, it is necessary to implement advanced signal processing and machine learning algorithms tailored to detect imperceptible changes in health parameters, aid early diagnosis and provide robust predictions.

This project aims to provide a broad set of computer-based methods capable of handling large amounts of multi-modal health-related data. Algorithms include tailor-made signal processing methods to improve signal quality and fuse data from different domains, as well as feature- and data-driven machine learning models that process data to provide a robust and comprehensive health assessment. The design, implementation and optimisation of prediction algorithms must consider the trade-off between performance, interpretability, and computational complexity. Indeed, algorithms capable of running in real time on autonomous, low-resource portable devices will be developed.

The research is inherently multidisciplinary and involves the application of computer methods to medical data. Data will be recorded in clinical settings under the supervision of medical personnel, as well as collected continuously in remote settings (e.g. at home). Young, healthy and elderly subjects as well as subjects with chronic and neurological diseases will be involved in the data acquisition process. Various commercial sensors and prototypes providing effective and non-invasive health monitoring will be exploited. Simple to complex algorithms will be developed and validated, both on large public data sets and on proprietary self-collected data. General models will be implemented to aid early diagnosis of chronic diseases and related symptoms. Subject-specific processing pipelines will be designed to match the characteristics of each patient, thus providing patient-centered solutions. Finally, secure and privacy-preserving methods, such as data anonymisation and federated learning, will be exploited to ensure adequate training of models while reducing the risk associated with the transfer of sensitive medical data.

The research results will be presented at recognised international scientific conferences (e.g., IEEE International Conference on Biomedical and Health Informatics, IEEE Symposium on Computer-Based Medical Systems, ACM International Symposium on Wearable Computers, IEEE International Conference on Digital Health) and top-tier international journals (e.g., Nature Digital Medicine, Artificial Intelligence in Medicine, Computers in Biology and Medicine, IEEE Journal of Biomedical and Health Informatics) at the intersection of computer science, engineering and medicine.

Research studies will be conducted and published in close cooperation with clinical professionals (e.g. neurologists, cardiologists, diabetologists, family doctors) and patient associations, to ensure that the methods developed meet current clinical gaps and the real needs of patients. In addition, national and international cooperation with academic research units (e.g. electronics, neuroscience, computer science) and specialised laboratories will provide a comprehensive set of knowledge and tools that will facilitate and promote high-quality research. Finally, collaborations with industrial partners will facilitate the design of effective, ready-to-use solutions for use in the real world.

Control theory-based sparse optimization for system identification  and data-driven control

The use of RNNs in the context of system identification and data-driven control poses a number of difficult problems in terms of efficiency of the training algorithms and complexity of the obtained models. In the last years, the research activity of the DAUIN system identification and control (SIC) group has been focusing on the application of control theory methods to design novel optimization algorithms that tackle the vanishing and exploding gradient problem affecting gradient-based training algorithms.
Although the obtained results are promising, a relevant problem still needs to be adequately addressed: how to force the training algorithm to learn models which are as simple as possible, to avoid redundancies, over-fitting, and undesired high complexity. This research line is drawing an increasing attention for its technological applications, such as, the implementation of the identified models and controllers in cyber-physical systems, sensor networks and smartphones. From a mathematical viewpoint, parsimonious models can be built by using sparse optimization tools, i.e., by solving suitable optimization problems where the number of effective parameters is encouraged to be small. For example, the sparsification can be obtained by suitable regularization methods.
The main goal of this Ph.D. project is to develop and analyze sparse optimization strategies to design suitable RNN models for system identification and data-driven contro purposes.

The research workplan is articulated as follows:

M1-M6: Study of the literature on the topics related to the research objective: sparse optimization, methods currently available to reduce the complexity of neural networks, recent results of the SIC group.

Milestone 1:
Report of the results available in the literature; theoretical formulation of problem and analysis of the main difficulties/critical aspect/open problems.

M7-M12: Development and analysis of a novel control theory -based sparse optimization technique for general convex problem.

Milestone 2:
Results obtained in this stage of the project are expected to be the core of a paper to be submitted to an international conference and/or an international journal.

M13-M24: Development and analysis of a novel control theory -based sparse optimization technique for general nonconvex problem. Application to training of RNNs.

Milestone 3:
Results obtained in this stage of the project are expected to be the core of a conference contribution and a paper to be submitted to an international journal.

M25-M36: Application of the proposed techniques to system identification and data-driven control of nonlinear systems. Theory and real-world application.

Milestone 4:
Application of the developed methods and algorithms to real-world problems, e.g., in the context of optimal control of electric vehicles and control of glucose-insuline dynamics in patients affected by Type I diabetes.

Main venues for journal publications: IEEE Transactions on Automatic Control, Automatica.

Main conferences: CDC, IFAC World Congress.

E-MIMIC:  Empowering Multilingual Inclusive Communication

Explainable AI (XAI) and Trustworthy AI algorithms for audio and speech data

The proposed PhD research will focus on developing novel explainable AI (XAI) techniques and trustworthy AI algorithms tailored explicitly for audio and speech data. The research activity will focus on the following targets.
-  Developing novel XAI techniques that provide insights into how AI models make decisions based on audio inputs.
- Ensuring the robustness and reliability of audio and speech AI models in real-world scenarios, particularly focusing on their ethical and unbiased performance.
- Validating the developed algorithms in practical tasks and applications such as automatic speech recognition, intent classification, and emotion recognition.
 
PhD years organization
YEAR I: the PhD candidate will review existing XAI and Trustworthy AI techniques for audio and speech data. The literature review will focus on by-design (e.g., concept-based) and post-hoc approaches (e.g., feature attribution-based explanations, counterfactual explanations). The candidate will define the research problem, identify gaps in the current state of the art, and formulate preliminary research hypotheses and opportunities for improvements. The candidate will then perform initial experiments to understand the limitations of current speech and audio models in terms of explainability and trustworthiness and explore preliminary solutions.
 
YEAR II: Based on the identified research gaps, opportunities for improvements, and the preliminary proposals made, the candidate will work on the design and implementation of novel XAI algorithms specifically for audio and speech data, ensuring their interpretability and transparency. The candidate will preliminarily evaluate the developed algorithms on benchmark datasets and real-world scenarios to refine their performance and reliability.
 
YEAR III: The candidate will consolidate the obtained findings. The candidates will focus on algorithm improvements, both in design and development, and applying the proposed novel approaches in multiple application domains. The candidate will conduct extensive evaluations to demonstrate the transparency, practical utility, and overall trustworthiness of the developed explainable algorithms and models.
 
The candidate's activities may include involvement with incoming projects, including a collaboration with Amazon, and potential industrial collaborations.
During the second/third year, the candidate will have the opportunity to spend a period abroad in a leading research center.

Publication venues for this research include leading conferences and journals in the fields of Explainable and trustworthy AI, audio and speech processing, machine learning , and artificial intelligence: 
IEEE/ACM TASLP (IEEE/ACM Transactions on Audio, Speech, and Language Processing)
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TIST (Trans. on Intelligent Systems and Technology)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Machine Learning with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
 
IEEE/ACM International Conferences (e.g., ACM KDD, IEEE ICDM, NeurIPS, ACM FAccT)
ACL Conferences (e.g., ACL, EMNLP, NAACL, EACL)
Audio and Speech processing Conferences (e.g., IEEE ICASSP, ISCA Interspeech)

Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) for the Insurance Industry

Large Language Models (LLMs) showcase impressive linguistic capabilities but produce hallucination, may incorporate outdated knowledge, and their predictions are non-transparent as the underlying ?reasoning? process is untraceable. Retrieval-Augmented Generation (RAG) has thus emerged as a promising solution by incorporating knowledge from external databases, including structured (knowledge graphs) and unstructured (textual sources). RAG has been shown to increase the accuracy and credibility of the generation, particularly for knowledge-intensive tasks, allowing for continuous knowledge updates and integration of domain-specific information. In particular, Knowledge Graphs (KGs) are structured representations that explicitly store rich factual knowledge. KGs can enhance LLMs by providing external knowledge for inference and interpretability. Meanwhile, KGs are difficult to construct and maintain over time; therefore, there is a need for automated methods in KGs to generate new facts and represent previously unseen knowledge. These limitations call for a synergistic approach in which the capabilities of both technologies are leveraged in a mutually beneficial way [1]. Starting from a core structured representation of domain-related concepts, entities and rules, such an approach could progressively and automatically expand the KG from unstructured textual documents. The KG could then provide a generalized structured representation of existing factual knowledge, that can in turn be used to enrich generated LLM-based queries, as well as verify the factual correctness of LLM-generated responses. 
 
The aim of the present PhD proposal is to exploire the integration and optimization of graph-enhanced RAG in LLMs to build robust and effective applications tailored to the insurance domain. To achieve this overarching goal, the following research questions will be tackled:- integrating information from KGs and other sources to perform RAG in LLMs for the insurance industry- designing LLM-based systems to perform KGs update, detecting novel concepts and inconsistencies- investigating to what extent RAG can improve the trusthworthiness of LLM-based applications in different contexts within the insurance industry 

Work plan outline- Review of the literature on RAG, KGs, and their integration with LLMs- Selection of one or more case studies- Definition of the requirements and specifications of the RAG-based solutions, along with suitable qualitative and quantitative evaluation methods- Design of novel methodologies for KGs embeddings in LLMs- Design of novel methodologies for LLM-based KG expansion and their integration into a continuous, life-long scenario- Implementation and evaluation of the proposed solutions- Dissemination of results in scientific journals and conferences
 
Possible publication venues: 
Results will be disseminated at conferences and international journals targeting both foundational and applied data science.  Conferences include European Conference on Machine Learning and Knowledge Discovery (ECML-PKDD), IEEE International Conference on Data Science and Advanced Analytics (DSAA), IEEE BigData, International Conference on Neuro-symbolic Learning and Reasoning, as well as emerging venues on generative AI and knowledge retrieval. Targeted journals include the IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Data and Knowledge Engineering, Pattern Recognition, Expert Systems with Applications. 
 
Industrial collaborations:
The activities will be conducted in collaboration with Reale Mutua Assicurazioni
 

[1] Pan, Shirui, et al. "Unifying large language models and knowledge graphs: A roadmap." IEEE Transactions on Knowledge and Data Engineering (2024).

Optimizing Deep Neural Networks on Edge SoCs using AI Compilers for Heterogenous Systems

Beyond TinyML: innovative Hardware Blocks and ISA Extensions for Next-Generation AI

XR-based Adaptive Procedural Learning Systems

In many fields, such as industry, medicine and others, procedures play an important role in the smooth execution of activities and tasks. In general, a procedure is a sequence of operations or activities that follow certain rules and ultimately serve to achieve a desired result. Procedural learning refers to the process of acquiring and mastering all the skills, knowledge and understanding required to perform a particular procedure effectively.
The main objectives of procedural learning can be summarized as follows. First, to provide learners with the necessary theoretical knowledge of the procedure, including an understanding of the underlying principles and concepts behind each step. Second, help learners acquire the practical skills they need to perform a procedure as accurately as possible, and to memorize and retain the individual steps of the procedure so that they can remember the sequence when they need to perform it. Finally, learners should also be enabled to master the procedure execution to minimize errors and optimize the workflow.
In the field of computer-assisted learning, XR (Extended Reality) technologies are certainly a game changer as they offer the possibility of providing multimedia and interactive information that utilizes contextual awareness, enableing the development of effective, engaging, and even entertaining learning environments.
This research project aims to advance the state of the art in XR-based procedural learning by tackling the following objectives:
-        Creating AI models that are able to adapt the learning content and its presentation based on the user's prior knowledge and progress. 
-        Development of XR environments that replicate different real-life scenarios with interactive elements and virtual instructors. 
-        Implement systems that provide real-time intelligent feedback tailored to learners' challenges and achievements, as well as advanced and natural interaction mechanisms (e.g. natural language interaction) to improve user engagement and the realism of training scenarios.
-        Implement autonomous, lifelike virtual agents that can take on different roles, from non-playable characters (NPCs) to virtual instructors. These agents will enhance the interactivity and realism of the training environment.
-        Develop solutions that are compatible with various XR platforms and scalable across different hardware capacities to ensure broad acceptance and future-proofing against technological advances.
-        Develop a framework for rapid prototyping and development of procedural learning applications featuring visual design tools for non-technical professionals to create training content and define activity graphsprocedures' activities, rules and constraints. This approach aims to improve collaborative development, simplify communication between technicians and domain experts, and automate the generation of software elements from visual representations.

Work Plan:
Phase 1 - Foundation and initial development (M1-M16)
Analyze the current SOTA in procedural learning, XR-based learning and adaptive learning systems and identify gaps and opportunities.
Create initial prototypes and conduct their preliminary testing and validation.

Phase 2 - XR procedural learning framework (M6-M30)
The framework will support portability and scalability across different devices and environments (from AR to VR) and will be iteratively extended with the following features:
- intelligent real-time feedback systems tailored to learners' challenges and achievements
- natural interaction mechanisms for better usability and user engagement
- autonomous, lifelike virtual agents that can assume different roles
- visual design tools for non-technical professionals and domain experts to create training content and define activity graphs, rules and constraints

Phase 3 - Refinement (M12-M36)
Identify one or more specific use cases and with the collaboration of domain expert design and implement a training activity exploiting the developed framework. Once developed, select a group of target learners and conduct comprehensive testing of the entire system. Collect data to asses user experience and learning effectiveness of the proposed ssyem. 
Refine the proposed systems according to the feedback gathered and iteratively improve the solution.

Publication Venues:
-   Journals: IEEE Transactions on Learning Technologies, Computers & Education, International Journal of Human-Computer Studies, Virtual Reality.
-   Conferences: CHI, IEEE VR, AIED, ICALT, ISMAR, EC-TEL.

Industrial Collaboration and Funding: collaborations with companies specializing in digital infrastructure for training, such as Immersave, and stakeholders like the MECAU department of the Ospedale Molinette of Turin, which has several ongoing collaboration with our group.
 

AI Based non-invasive vital signs monitoring

The scope of the research is to study and develop a robust methodology to monitor a driver in vehicle by detecting continuously vital parameters such Heart Rate and SpO2. The invasiveness of the device must be reduced to the maximum possible extent, so as to not to affect the status of the driver him/herself. In principle, non-contact monitoring would be the preferred solution. To this end, devices like global shutter cameras will be explored with different illumination system in the IR domain in order not to blind the driver. Moreover, a robust methodology for signal processing, face detection, recognition of the points of acquisition and reduction of the noise introduced by vibration or human motion will be devised and experimented.
Deep or machine learning methodologies will be evaluated to reduce motion artifacts, focusing both on performance and real time response with very short latency.
In parallel of the main activity, a collection of ECGs will be performed to improve the automatic identification of cardiac anomalies. 

WORKPLAN: YEAR 1.
Task 0. Revision of the state of the art related to in-vehicle vital signs remote monitoring. 
Task 1. Collection of ECG data in hospital and RSA to set up algorithms for anomaly detection (contact ECG sensors).
Task 2: Implementation and testing if algorithms for hearth rate (HR) anomaly detection and classification (contact ECG sensors).
Task 3. Development of a prototypal system for contactless PPG detection in vehicle.
YEAR 2.
Task 4. ML and DL models to clean up the signal from electrical and background noise.  Implementation and testing of algorithms for HR anomaly detection and classification (contactless).
Task 5. Collection of databases in vehicle to train the models.
Task 6. First results of the prototypal system in vehicle. 
YEAR 3.
Task 7. Finalization of the data collection champaign in hospital, and related data analysis. 
Task 8. ML and DL models to reduce the noise due to the motion and vibration.
Task 9. Finalization of the algorithms and the preparation of  demo in vehicle during normal usage
Task 9. Critical analysis of the results, proposal of integration of the algorithms with a robust trial on the field.
Task 10. Critical analysis of the results, proposal of integration of the algorithms with heterogeneous data present in the patient's clinical records.

We plan to have at least two journal papers published per year.
Target journals:
IEEE Transactions on Biomedical Engineering
IEEE Journal on Biomedical and Health Informatics
IEEE Access
IEEE Journal of Translational Engineering in Health and Medicine
MDPI Sensors
Frontiers in Neurology
 
COOPERATIONS:
-CNR-IEIIT
-Mauriziano Hospital,  Cardiology Department

A cooperation is on going with Cooperativa Valdocco to perform measurements of cardiac activity of elderly persons, using wearable devices.

ON GOING PROJECTS ON RELATED TOPICS:
-        PNRR Ecosistema NODES Nord-Ovest Digitale e Sostenibile
-        PNRR Salute Complementare "Health Digital Driven Diagnostics, prognostics and therapeutics for sustainable Health care"
-        PRIN-PNRR 2022 "Objective monitoring of axial symptoms in Parkinson's disease: quantitative assessment in daily life based on the use of wearables, video sensing and artificial intelligence (OMNIA-PARK)"
 

Statistical learning of genomic data

High-End AI-Enhanced Computing System for Space Application

The research focuses on methodologies to enhance the reliability, efficiency, and capabilities of AI-driven systems in space to support the critical role of AI in space applications outlined above. 

The main challenges in developing high-end AI-enhanced computing systems for space applications include:Reliability of the systemsReal-time ProcessingLimited Computational Resources

Artificial Intelligence (AI) is increasing its role in space applications, revolutionizing various aspects of space exploration, satellite operations, and data analysis. AI algorithms have proven extremely effective in enabling autonomous navigation and driving for spacecraft and remote vehicles. For example, NASA's Mars rovers, such as Curiosity and Perseverance, use AI to traverse the Martian terrain, avoiding obstacles and selecting paths to scientific targets. 
 
AI also monitors the health and status of satellites, predicting potential failures and optimizing maintenance schedules. It is used for processing and analyzing images captured by Earth observation satellites, supporting applications in environmental monitoring, disaster response, and agriculture. Such approaches can significantly enhance Earth observation and space image analysis, bringing new capabilities and efficiencies to these applications. AI models analyze satellite data to monitor indicators of climate change, such as melting ice caps, sea-level rise, and changes in vegetation cover. Using AI can perform data analysis on edge to reduce satellite data transfer tasks.
 
Typically, AI-driven functionalities rely on embedded systems for real-time data processing and decision-making. Embedded systems equipped with AI capabilities allow satellites and spacecraft to process data onboard, significantly reducing the need for data transmission to Earth and enabling real-time analysis and decision-making. For instance, CubeSats and larger satellites often include embedded systems that use AI algorithms to process imagery and sensor data immediately after collection. System-on-chip (SoC) platforms integrate processors, memory, and AI accelerators into a single chip, providing the computational power required for AI tasks while minimizing space, weight, and power consumption, which represent critical factors in space-borne systems.

The research focuses on methodologies to enhance the reliability, efficiency, and capabilities of AI-driven systems in space to support the critical role of AI in space applications outlined above. 
 
The main challenges in developing high-end AI-enhanced computing systems for space applications include:
 - Reliability of the systems: High levels of radiation characterizing the space environment are a primary source of errors in space applications that must be addressed at the hardware and application level. Advanced research in continuous availability, fault-tolerant designs, and redundancy strategies are fundamental to exploring and enhancing the reliability of AI-driven systems.
 - Real-time Processing: Space applications often require real-time data processing for navigation, hazard avoidance, and scientific analysis. The research aims to develop AI systems capable of high-speed, real-time processing with minimal latency, exploiting the latest devices and architecture, focusing particularly on heterogeneous and parallel paradigms for edge devices.
 - Limited Computational Resources: Systems on the edge often have constrained computational resources due to size, weight, and power limitations. Developing AI algorithms that can operate efficiently within these constraints is a non-trivial challenge. The research must explore lightweight, resource-efficient AI models that deliver high performance relying on the available computational infrastructure. This includes a high codesign of a system based on techniques such as model pruning, quantization, and customized AI accelerators.
 
 
The research projects correlated to the present proposals in the last years are:
-      [active 2023] ?TERRAC: Towards Exascale Reconfigurable and Rad-hard Accelerated Computing in space? in cooperation with the European Space Agency (ESA)
-      [close 2020] ?The HERA Cubesat program? in cooperation with DIMEAS, Tyvak International and the European Space Agency (ESA)
 
The research program will target the following conference venues: 
-       IEEE Design and Test in Europe (DATE) 
-       IEEE Dependable Network and System (DSN)
-       IEEE International Test Conference (ITC)
-       IEEE International Conference on Field Programmable Logic (FPL)
-       IEEE RADECS
-       IEEE NSREC
-       ACM Computing Frontiers
 
The research program will target the following journals: 
-       IEEE Transactions on Circuits and Systems-I
-       IEEE Transactions on Computers
-       IEEE Transactions on Reliability
-       ACM Transactions on Reconfigurable Technology and Systems
-       Intelligent Systems with Applications 
-       Transactions of Emerging Topics in Computing

Design and development of software stack for execution of hybrid Quantum-classic algorithms

Quantum computing is a very powerful technology whose purpose is to speed-up traditional computation, given its exponential intrinsic parallelism, able to make linear the exponential algorithms.
However, given the reduced number of qubits available today, NISQ algorithms are usually hybrid.
For this reason a strict integration with classic resources is fundamental to be able to work together with technologies like cloud computing and High Performance Computing (HPC).
PhD candidate will work on the definition of an open source software stack for superconducting quantum machines, in order to integrate them with classic computers for solving scientific and industrial problems.
In particular, activity will be concentrated on the research of state of the art solutions, on implementing new software components and on the integration with vendors of quantum hardware for the following abstraction levels:- controlling the machine by using pulses, integrating standard code with pulse-level code - compiling efficiently (exact or approximated compiling) quantum circuits for a given hardware target architecture, defining benchmarks for understanding the effect of compiling on potential applications - integrating quantum architecture with standard architecture, simplifying access to the classic counterpart, and efficiently allocating quantum resources together classic resources 
The proposal puts together competences of Dipartimento di Automatica e Informatica and knowledge of Fondazione Links. Experimental activities will be done mainly together Fondazione Links.
This work will be developed during the three years, following the usual Ph.D program:
-       first year, improvement of the basic knowledge about quantum computing, attendance of most of the required courses, also on quantum mechanics if required, submission of at least one conference paper
-       second year, design and implementation of new algorithms for the previously illustrated abstraction levels and submission of conference papers and at least one journal
-       third year, finalization of the work, with at least a selected journal publication.

Possible venues for publication will be, if possible, journals and conferences related to HPC and Quantum Computing, from IEEE, ACM and SPIE. An example could be the IEEE Transactions on Quantum Engineering.
The scholarship is sponsored by Fondazione Links.  A period of six months abroad will be done during the PhD, and a period of at least six months in Fondazione Links will be mandatory too.
The work will therefore be done in strict collaboration together with Fondazione Links, with whom there is already a long-term collaboration, not only on QC.
 

RESEARCH OBJECTIVES

1. Build a comprehensive framework for the existing and potential synergies between SE and AI. The objective is to explore, document and organize in a conceptual framework the complementary aspects of SE and AI, deeply analyzing and identifying the areas within SE that can benefit from the effective application of AI techniques and vice versa.

2. Develop Customized SE Techniques for AI Lifecycle.
This objective focuses on developing tailored SE techniques to improve the AI lifecycle in a set of selected dimensions such as reliability, robustness, and interpretability of AI systems, and according to specific stages (es., dataset collection and management, feature engineering, model training, model deployment). The high-level goal is to ensure AI systems align with established SE principles and standards.

3. AI-Driven Approaches for Software Quality Assurance
This objective focuses on the creation of novel AI techniques specifically tailored for software quality assurance of SE processes (e.g., requirements elicitation and verification, code review, software testing). The general goal is to fortify the reliability and robustness of software systems through AI-powered methodologies.

OUTLINE OF THE WORKPLAN

1) Literature review and framework development
Conduct an extensive review of the existing literature on Artificial Intelligence for Software Engineering (AI4SE) and Software Engineering for Artificial Intelligence (SE4AI). Identify current proposals and implementations, research gaps and potential areas for further investigation. The findings will be organized within a comprehensive conceptual framework, that should consider both the specific phases of software development and the stages of a typical AI pipeline. This will take place mainly during the 1st year.

2) Development/customization and evaluation of SE Techniques for AI improvement.  Building on top of the previous activity's findings, the candidate shall select a set of SE techniques to be specifically tailored for the AI lifecycle. Examples are design methodologies for requirements elicitation, testing strategies, maintenance techniques and tools. Experimental evaluation should validate the improvements in terms of more robust and reliable AI systems, comparing the outcomes achieved with and without SE techniques. Comparison with existing benchmarks is also possible. This will take place mainly during the 2nd and 3rd  years.

3) Development/customization and evaluation of AI-Driven Approaches for software quality assurance. Based on the findings of the first activity, the candidate will develop or customize AI techniques to address the identified SE quality assurance challenges, either in terms of processes or artifacts quality. This activity may involve using one or more of the following techniques: ML algorithms, neural networks, natural language processing, image recognition/classification techniques. Experimental evaluation should validate the improvements in the involved aspects of the software development lifecycle, comparing the performance and effectiveness of the AI-driven novel approach with existing SE methods. This should allow to highlight the advantages, disadvantages, and potential improvements offered by the AI-driven approaches. This will take place mainly during the 2nd and 3rd  years.

LIST OF POSSIBLE VENUES FOR PUBLICATION

Journals:
- IEEE Transactions on Software Engineering
- Empirical Software Engineering 
- ACM Transactions on Information Systems
- European Journal of Information Systems
- Journal of Systems and Software 

Conferences:
- ACM/IEEE International Symposium on Empirical Software Engineering and Measurement
- International Conference on Software Engineering
- International Conference on Product-Focused Software Process Improvement
- International Conference on Evaluation and Assessment in Software Engineering (EASE) 

The candidate should exhibit a good mix of: 
- Solid foundation in SE principles, methodologies, and practices. He/she should be familiar with the software development lifecycle and software quality assurance techniques.
- Good understanding of AI techniques. He/she should be proficient in the use of ML and AI techniques, with evidence of past project experiences.

He/she should also be able to communicate the research in an effective way.

Protect-IT | Privacy on the Internet: measurements and novel approaches

Research objectives:

In the current web ecosystem, targeted or behavioural advertising lets providers monetize their content, by collecting and processing personal data to build accurate user profiles. 
This massive data collection has created tension between users and the ads ecosystem. Mozilla Firefox and Apple Safari have started battling third-party cookies by giving third-party cookies a separate cookie jar per site, so they cannot be used to track users across sites anymore. Other alternatives have been proposed to limit and control the amount of information users share with web services.

This project consists of the design, implementation and testing of novel measurement methodologies to observe and quantify the amount of information online services collect, and on the definition of possible alternative solutions that result in more privacy friendly.

The research will focus on the design of data collection methodologies that allow one to observe and quantify the amount of information web services collect when users browse the web. This includes both active measurements, e.g., web crawlers, and passive measurements, e.g., browser plugins that observe the users' browsing activity. We will investigate mechanisms that will offer privacy guarantees to users, e.g., via the extraction of client-side models that process the information locally rather than sharing the raw data with servers. At the same time, we will investigate the introduction of privacy guarantees, e.g., based on differential privacy or k-anonymity concepts.

In the first phase, the candidate will study the state of the art of privacy methodologies in general and of the problem of data collection on the internet. He/she will design possible data collection methodologies and metrics to measure and quantify the current state.

In the second phase, the candidate will design and integrate possible alternative solutions to mitigate and control the amount of information web services collect from users. These shall include privacy guarantees for example those provided by differential privacy or k-anonymity approaches.

The project will involve a collaboration with partners in the Protect-IT project including the University of Brescia, of Naples, of Milano among others.

Outline of the research work plan:

1st year
- Study of the state-of-the-art on privacy and the web, initial design of the data collection platforms and machine learning models to extract and quantify the amount of personal data exchanged with web services.

2nd year
- Data collection campaign and measurement analysis
- Propose and develop innovative solutions to the problems of privacy on the web.
- Propose machine learning approaches to offer privacy guarantees and extract local models to avoid the sharing of raw data.

3rd year 
- Tune the developed techniques and highlight possible strategies to deploy them in the wild.
- Application of the strategies to new data for validation and testing.

References:
-    Nikhil Jha, Martino Trevisan, Luca Vassio, and Marco Mellia. 2022. The Internet with Privacy Policies: Measuring The Web Upon Consent. ACM Trans. Web 16, 3, Article 15 (August 2022), 24 pages. https://doi.org/10.1145/3555352
-    N. Jha, M. Trevisan, M. Mellia, R. Irarrazaval and D. Fernandez, "I Refuse if You Let Me: Studying User Behavior with Privacy Banners at Scale," 2023 7th Network Traffic Measurement and Analysis Conference (TMA), Naples, Italy, 2023, pp. 1-9, doi: 10.23919/TMA58422.2023.10198936. 
-    Nikhil Jha, Martino Trevisan, Emilio Leonardi, and Marco Mellia. 2024. Re-Identification Attacks against the Topics API. ACM Trans. Web Just Accepted (June 2024). https://doi.org/10.1145/3675400

List of possible venues for publications_

Security venues: IEEE Symposium on Security and Privacy, IEEE Transactions on Information Forensics and Security, ACM Symposium on Computer and Communications Security (CCS), USENIX Security Symposium, IEEE Security & Privacy;

AI venues: Neural Information Processing Systems (NeurIPS), International Conference on Learning Representations (ICLR), International Conference on Machine Learning (ICML), AAAI Conference on Artificial Intelligence, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD);

Computer networks venues: Distributed System Security Symposium (NDSS), Privacy Enhancing Technologies Symposium, The Web Conference (formerly International World Wide Web Conference WWW), ACM International Conference on Emerging Networking EXperiments and Technologies (CoNEXT), ACM Internet Measurement Conference, USENIX Symposium on Networked Systems Design and Implementation (NSDI).

Protect-IT | Distributed platform for cybersecurity data collection and analysis

Research objectives:

The project consists of the design, implementation and testing of a flexible and distributed infrastructure to collect and process data related to the identification of novel cybersecurity threats. Thanks to technologies based on containers, darknets, and honeypots, the infrastructure will support the collection of data in a scalable manner. The usage of AI and ML solutions will allow one to extract actionable information to identify already known and novel attacks. To avoid data sharing, Federated Learning solutions will allow the training of a common model by leveraging the data each vantage point will collect seamlessly.

The research will focus on extending machine learning solutions to automate and assist security analysts in the process (i) of identifying new attacks and (ii) setting up countermeasures.

In the first phase, the candidate will study the state of art of platforms for cybersecurity data collection and machine learning, including the problem of data collection in distributed scenarios, and of federated learning solutions.

In the second phase, the candidate will create and setup the distributed platform leveraging servers and virtual machines offered by partners of the PROTECT-IT/SERICS project. This will include the joint deployment of (i) data collected from honeypots and (ii) darknets.
Then he/she will investigate machine learning models that can identify new threats. For this, he/she will leverage mechanisms to create generic data representation using embeddings that will be then specialised to solve custom downstream tasks. Via federated learning, the candidate will investigate how to train a shared embedding model suitable to then solve specific tasks such as threat classification, anomaly detection, or identifying new attack patterns.

The project will involve a collaboration with partners in the Protect-IT project including the University of Brescia, Naples, Milano among others.

Outline of the research work plan:

1st year
- Study of the state-of-the-art security log analysis and state-of-the-art data collection platforms and machine learning models in ML.
- Data collection platform design with inclusion of security monitors such as honeypots, darknets, IDS, etc.

2nd year
- Deployment of the distributed platform for data collection and initial model training using federated learning.
- Propose and develop innovative solutions to the problems of cyber threats analysis with machine learning solutions.
- Propose multi-modal embeddings for network raw data and security logs.

3rd year 
- Tune the developed techniques and highlight possible strategies to counteract the various threats.
- Application of the strategies to new data for validation and testing.

References:

-    Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Drago, Idilio, Milan, Giulia, Houidi, Zied Ben, Rossi, Dario (2023). Cross-network Embeddings Transfer for Traffic Analysis. IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, p. 1-13, ISSN: 1932-4537, doi: 10.1109/TNSM.2023.3329442
-    Luca Gioacchini, Luca Vassio, Marco Mellia, Idilio Drago, Zied Ben Houidi, Dario Rossi (2023). i-DarkVec: Incremental Embeddings for Darknet Traffic Analysis. ACM TRANSACTIONS ON INTERNET TECHNOLOGY, vol. 23, p. 1-28, ISSN: 1533-5399, doi: 10.1145/3595378
-    Huang, Kai, Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Dynamic Cluster Analysis to Detect and Track Novelty in Network Telescopes, 9th International Workshop on Traffic Measurements for Cybersecurity (WTMC 2024), Vienna, Austria, July 2024
-    Huang, Kai, Gioacchini, Luca, Mellia, Marco, Vassio, Luca, Incremental Federated Host Embeddings for Network Telescopes Traffic Analysis, IEEE International Workshop on Generative, Incremental, Adversarial, Explainable AI/ML in Distributed Computing Systems (AI-DCS), Jersey City, New Jersey, USA, July 2024

List of possible venues for publications:

Security venues: IEEE Symposium on Security and Privacy, IEEE Transactions on Information Forensics and Security, ACM Symposium on Computer and Communications Security (CCS), USENIX Security Symposium, IEEE Security & Privacy;

AI venues: Neural Information Processing Systems (NeurIPS), International Conference on Learning Representations (ICLR), International Conference on Machine Learning (ICML), AAAI Conference on Artificial Intelligence, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD);

Computer networks venues: Distributed System Security Symposium (NDSS), Privacy Enhancing Technologies Symposium, The Web Conference (formerly International World Wide Web Conference WWW), ACM International Conference on Emerging Networking EXperiments and Technologies (CoNEXT), USENIX Symposium on Networked Systems Design and Implementation (NSDI).

Human motion reconstruction and recognition based on dynamic point clouds

Research objectives

In recent years, works like that by Mei Oyama et al. [2] showed how to recover an accurate human pose from a single depth map. Pose alone, however, is not enough to provide useful information in the scenarios of interest. Dynamic point cloud data need to be considered for reconstructing the entire motion process. For motion recognition, Biswas et al. [3] designed an end-to-end system that combines RGB images and point cloud information to recover 3D human motion. Arindam Sengupta et al. [4], in turn, proposed a method of real-time detection and tracking of human joints using mmWave radars radar. These methods, though, infer motions based on the current pose, which introduces uncertainty. Moreover, it is challenging to use mmWave radars for object tracking, as their signals are usually of low resolution and highly noisy due to multi-path and specularity [5].

To cope with the above limitations, this research will directly consider a continuous sequence of mmWave point clouds over time to determine the motion. Moreover, to record the continuous changes in the point cloud, tracking algorithms will be applied to human body joints. Finally, in order to cope with challenges associated with pure radar-based reconstructions, the combination of vision and radar will be explored to support applications in non-ideal configurations.

Outline of the proposed project plan

During the first year, the PhD student will review the state of the art regarding multi-modal 3D reconstruction, detection and recognition of human motion. He or she will start familiarizing with open problems in domains encompassing, e.g., gaming, movie and game production, surveillance, home and industry automation, health monitoring, etc. by focusing on one of them and leveraging mmWave radar technology alone first. At least a conference publication is expected to be produced based on the results of these activities. 
In the second year, the student will work on extending the scope of the analysis by addressing the limitations found and devising solutions based on the fusion of the two types of data (mmWave radar and depth camera). The student is expected to publish at least another conference paper on the outcomes of these activities. 
During the first two years, the student will complete his or her background in topics relevant for the proposal by attending suitable courses. 
During the third year, the student will work onto evaluating the effectiveness of the devised solutions in realistic scenarios by also constructing appropriate datasets involving human users. Results obtained will be reported into at least a high-quality journal publication.  

Possible publication venues

Publications will target conferences and journals in the areas of multi-modal machine/deep learning, pattern recognition, and human-computer interaction like, e.g., IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Human-Machine Systems, ACM Transactions on Computer-Human Interaction, ACM Conference on Human Factors in Computing Systems (CHI), CVPR, ICPR, etc. Some specific publication venues related to the considered application scenarios could be considered too.

References

[1] J. Yang, A. W. Ismail , "A review: Deep learning for 3D reconstruction of human motion detection," in International Journal of Innovative Computing, 12(1), 65?71, 2021.
[2] M. Oyama, N. Kaneko, M. Hayashi, K. Sumi and T. Yoshida, "Two-stage model fitting approach for human body shape estimation from a single depth image," Proc. 15th IAPR International Conference on Machine Vision Applications (MVA), 2017.
[3] A. Biswas, H. Admoni, A.Steinfeld, "Fast on-board 3D torso pose recovery and forecasting ," Proc.  International Conference on Robotics and Automation, 2019.
[4] A. Sengupta, F. Jin, R. Zhang, S. Cao, "mm-Pose: Real-time human skeletal posture estimation using mmWave radars and CNNs," in IEEE Sensors Journal, vol. 20, no. 17, pp. 10032-10044, 2020.
[5] Y. Sun, Z. Huang, H. Zhang, Z. Cao, D. Xu, "3DRIMR: 3D reconstruction and imaging via mmWave radar based on deep learning," IEEE International Performance, Computing, and Communications Conference (IPCCC), 2021.

Data-Driven and Sustainable Solutions for Distributed Systems

Two research questions (RQ) guide the proposed work:

RQ1: How can we design and implement on local and larger-scale testbeds effective autonomous solutions that integrate the network information at different scopes using recent advances in supervised and reinforcement learning?

RQ2: To scale the use of machine learning-based solutions in cyber-physical systems, what are the most efficient distributed machine learning architectures that can be implemented at the edge of such systems?

The final target of the research work is to answer these questions, also by evaluating the proposed solutions on small-scale emulators or large-scale virtual testbeds, using a few applications, including virtual and augmented reality, precision agriculture, or haptic wearables. In essence, the main goals are to provide innovation in decision, planning, responsiveness, using centralized and distributed learning integrated with edge computing infrastructures. Both vertical and horizontal integration will be considered. By vertical integration, we mean considering learning problems that integrate states across hardware and software, as well as states across the network stack across different scopes. For example, the candidate will design data-driven algorithms for planning the deployment of IoT sensors, tasks scheduling, and resources organization. By horizontal learning, we mean using states from local (e.g., physical layer) and wide area (e.g., transport layer) as input for the learning-based algorithms. The data needed by these algorithms are carried to the learning actor by means of newly networking protocols.  Aside from supporting resiliency with the vertical integration, solutions must offer resiliency across a wide (horizontal) range of network operations:  from close-edge, i.e., near the device, to the far-edge, with the design of secure data-centric resource allocation (federated) algorithms.

The research activity will be organized in three phases:

Phase 1 (1st year): the candidate will analyze the state-of-the-art solutions for cyber-physical systems management, with particular emphasis on knowledge-based network automation techniques. The candidate will then define detailed guidelines for the development of architectures and protocols that are suitable for automatic operation and (re-)configuration of such deployments, with particular reference to edge infrastructures. Specific use-cases will also be defined during this phase (e.g., in virtual reality, smart agriculture). Such use cases will help identifying ad-hoc requirements and will include peculiarities of specific environments. With these use cases in mind, the candidate will also design and implement novel solutions to deal with the partial availability of data within distributed edge infrastructures. Results of this work will likely result in conference publications.

Phase 2 (2nd year): the candidate will consolidate the approaches proposed in the previous year, focusing on the design and implementation of mechanisms for vertical and horizontal integration of supervised and reinforcement learning. Network, and computational resources will be considered for the definition of proper allocation algorithms, with the objective of energy efficiency. All solutions will be implemented and tested. Results will be published, targeting at least one journal publication.

Phase 3 (3rd year): the consolidation and the experimentation of the proposed approach will be completed. Particular emphasis will be given to the identified use cases, properly tuning the developed solutions to real scenarios. Major importance will be given to the quality offered to the service, with specific emphasis on the minimization of latencies in order to enable a real-time network automation for critical environments (e.g., telehealth systems, precision agriculture, or haptic wearables). Further conference and journal publications are expected.

The research activity is in collaboration with Saint Louis University, MO, USA and University of Kentucky, KY, USA, also in the context of some NSF grants.

The contributions produced by the proposed research can be published in conferences and journals belonging to the areas of networking and machine learning (e.g. IEEE INFOCOM, ICML, ACM/IEEE Transactions on Networking, or IEEE Transactions on Network and Service Management) and cloud/fog computing (e.g. IEEE/ACM SEC, IEEE ICFEC, IEEE Transactions on Cloud Computing), as well as in publications related to the specific areas that could benefit from the proposed solutions (e.g., IEEE PerCom, ACM MobiCom, IEEE Transactions on Industrial Informatics, IEEE Transactions on Vehicular Technology).

Multi-Provider Cloud-Edge Continuum

Cloud-native technologies are increasingly deployed at the edge of the network, usually through tiny datacenters made by a few servers that maintain the main characteristics (powerful CPUs, high-speed network) of the well-known cloud datacenters. This project proposes to aggregate the above available hardware into a gigantic, distributed datacenter, potentially controlled by multiple administrative entities. This would originate a set of highly dynamic, yet still secure and robust, ?virtual? clusters spanning across multiple physical infrastructures. 

The key objectives of the present research are the following:
- Architectural Paradigms and Resource Management: Develop new paradigms, algorithms, and protocols for advertising, negotiating, and acquiring resources across different administrative domains.
- Orchestration and Lifecycle Management: Create scalable algorithms and protocols for orchestrating the computing continuum, optimizing the placement of software services, and managing the lifecycle of virtual clusters.
- Resilience and Evolution: Ensure the software infrastructure and virtual clusters can survive and evolve during network outages or planned disconnections.
- Data Exchange Optimization: Optimize data exchange between virtual services when clusters span multiple sites.

The research activity is part of the IPCEI projects initiative (https://www.bmwk.de/Redaktion/EN/Artikel/Industry/ipcei-cis.html), in particular in collaboration with TIM S.p.A..

The research activity will be organized in the following phases.

Year 1: Foundation and Initial Development
Q1-Q2: Literature Review and Requirement Analysis
-  Conduct a comprehensive literature review on cloud-native technologies, edge computing, and distributed datacenters.
-  Identify key challenges and gaps in current technologies.
-  Define detailed requirements for the project.
Q3-Q4: Architectural Design and Initial Prototyping
-  Develop novel architectural paradigms for resource management across multiple administrative domains.
-  Design initial scalable algorithms and protocols for resource advertisement, negotiation, and acquisition.
-  Create a prototype for testing basic functionalities.

Year 2: Development and Testing
Q1-Q2: Advanced Algorithm Development
-  Develop scalable algorithms and protocols for orchestrating the computing continuum.
-  Focus on optimized placement for software services and applications.
-  Implement lifecycle management for virtual clusters.
Q3-Q4: Resilience and Data Exchange Optimization
-  Design and implement algorithms for infrastructure resilience during network outages or planned disconnections.
-  Develop scalable algorithms for optimizing data exchange between virtual services across multiple sites.
-  Conduct extensive testing and validation of the developed algorithms and protocols.

Year 3: Integration and Evaluation
Q1-Q2: Integration and System Testing
-  Integrate all developed components into a cohesive system.
-  Perform comprehensive system testing to ensure functionality and performance.
-  Optimize the system based on testing results.
Q3-Q4: Evaluation and Dissemination
-  Evaluate the system in real-world scenarios to assess its effectiveness and robustness.
-  Document findings and prepare research papers for publication.
-  Present results at conferences and workshops.
-  Plan for potential exploitation or further research extensions.

Expected target conferences are the following:
Top conferences:
- USENIX Symposium on Operating Systems Design and Implementation (OSDI)
- USENIX Symposium on Networked Systems Design and Implementation (NSDI)
- International Conference on Computer Communications (INFOCOM)
- ACM European Conference on Computer Systems (EuroSys)
- ACM Symposium on Principles of Distributed Computing (PODC)
- ACM Symposium on Operating Systems Principles (SOSP)

Journals:
- IEEE/ACM Transactions on Networking
- IEEE Transactions on Computers
- ACM Transactions on Computer Systems (TOCS)
- IEEE Transactions on Cloud Computing

Magazines:
- IEEE Computer
- IEEE Network

Reliability estimation techniques for AI accelerators

Analysis and improvement of fault tolerance and safety of AI accelerators

Multimodal Spatio-Temporal Data Analytics

The main objective of this research activity is to design data-driven techniques and machine learning algorithms to fuse and analyze contemporaneously multimodal Spatio-Temporal data (e.g., numerical time series collected by sensors combined with sequences of satellite images or tabular data with temporal and spatial information). Both descriptive and predictive problems will be considered.

The main research objectives that will be addressed are as follows.

Multimodality of the sources. A plethora of spatio-temporal data sources associated with different modalities are available. Each source and modality provides a different perspective of the phenomena under analysis. However, integrating different spatio-temporal sources and modalities is still not well-managed. Effectively combining different sources and modalities allows the mining of new knowledge that cannot be extracted considering one source or modality at a time. Innovative integration and fusion techniques based, for instance, on deep neural networks will be studied to address this challenge.

Big spatio-temporal data. Spatio-Temporal data are frequently big (e.g., vast collections of sequences of remote sensing data) or are gathered using big data streaming frameworks. Hence, specific multimodal big data systems and approaches are needed to collect, store, and manage spatio-temporal data.

Resource awareness. In several domains, near-real-time analyses are needed. In this context, part of the analysis is performed onboard the devices used to collect the data. Resource-aware machine learning models are needed in that case to use the limited amount of resources as best as possible.

The work plan for the three years is organized as follows.

1st year. Analysis of the state-of-the-art algorithms and data analytics frameworks for multimodal Spatio-Temporal data analysis. Based on the pros and cons of the current solutions, a preliminary approach for integrating spatio-temporal sources with different modalities will be studied and designed. Based on the proposed data integration approaches, novel algorithms will be designed, developed, and validated on historical spatio-temporal data related to specific domains (e.g., water level monitoring of lakes).

2nd year. The scalability of the proposed models will be mainly addressed during the second year. Starting from the algorithms designed during the first year, big data versions of the algorithms will be developed and validated. Moreover, the proposed algorithm will be evaluated in different domains.

3rd year. Resource-aware models will be designed for onboard inference to analyze in near-real-time and onboard the collected spatio-temporal data.

The outcomes of the research activity are expected to be published at IEEE/ACM International Conferences and in any of the following journals:
- ACM Transactions on Spatial Algorithms and Systems
- IEEE Transactions on Knowledge and Data Engineering
- IEEE Transactions on Big Data
- IEEE Transactions on Emerging Topics in Computing
- ACM Transactions on Knowledge Discovery in Data
- Information sciences (Elsevier)
- Expert systems with Applications (Elsevier)
- Machine Learning with Applications (Elsevier)

Reinforcement Learning for Multi-Scale Energy Management in Multi-Energy Systems

4D reconstruction and perception

Video Dynamic Scene Understanding

Heterogeneous Computing Design Techniques for HPC

This research proposal aims to explore and develop strategies for leveraging heterogeneous computing architectures to enhance the performance and efficiency of artificial intelligence (AI) and machine learning (ML) workloads.
The main objectives of the research proposal are: Investigate Workload Partitioning Techniques: Develop methods for dividing AI/ML tasks into smaller subtasks that can be processed concurrently by different types of state-of-the-art processors and co-processors. This stage will require exploring algorithms for dynamically allocating tasks to processors based on real-time load and availability, maximizing resource utilization, and minimizing bottlenecks. This objective is expected to be achieved during the first year of the PhD program. Design of custom HPC node on Reconfigurable SoC: With acquired knowledge of AI computational paradigms, this stage aims to explore different architectural solutions implementing custom AI accelerator with reconfiguration capabilities. The new architectural solution should adapt its datapath dynamically depending on the workload execution to manage the available resources while providing power efficiency efficiently. The custom HPC node will be implemented by exploiting high-performance reconfigurable SoC equipped with a next-generation FPGA device. This objective is expected to be achieved within the second year of the PhD program. Validation and Performance Exploration: The developed HPC  node will be validated by developing a monitoring system. A host PC software platform will be developed to generate and transmit the test workload. The developed accelerator performance capabilities should be compared with current state-of-the-art solutions. For this purpose, a specific software compiler should be implemented to automatize the process of translating high-level algorithm models to the specific Instruction Set Architecture (ISA) of the developed accelerator. This objective is expected to be achieved within the third year of the PhD program. Inclusion of the custom Accelerator in HPC system: The custom reconfigurable AI accelerator will be included in a Heterogenous HPC system consisting of multiple CPUs, GPUs and FPGA devices. At this stage, the focus will be on selecting the optimal transmission protocol to enable data transfer among the different COTS units and the custom HPC node, while also addressing different nodes' latency. Particular emphasis will be placed on investigating the use of shared memory spaces accessible by all nodes to reduce data transfer overhead and improve overall efficiency. Moreover, it will be required to create software frameworks that facilitate uniform communication and data transfer between the heterogeneous nodes, ensuring optimal workload distribution. This objective is expected to be achieved within the third year of the PhD program.

The research projects correlated to the present proposals in the last years are:[active 2024] ?AMATERASU: Analysis, Mitigation, And Radiation Test of Advanced COTS FPGAs through Reliability-Aware Place and Route? financed by the European Space Agency (ESA)- [active 2023] ?REASE: Resilient computing Architecture in the Space Quantum Communication Era? ? PRIN 2023- [active 2023] ?TERRAC: Towards Exascale Reconfigurable and Rad-hard Accelerated Computing in space? financed by the European Space Agency (ESA)- [active 2021] ?EuFRATE: European FPGA Radiation-hardened Architecture for Telecommunications and Embedded computing in space? in cooperation with Argotec and the European Space Agency (ESA)- [active 2020] ?The HERA Cubesat program? in cooperation with DIMEAS, Tyvak International and the European Space Agency (ESA)

The research program will target the following conference venues: - IEEE Design and Test in Europe (DATE) - IEEE Conference on Computer Design (ICCD)- IEEE International Symposium on Circuits and Systems (ISCAS)- IEEE International Conference on Field Programmable Logic (FPL)- ACM International Conference on Computing Frontiers (CF)
The research program will target the following journals: - IEEE Transactions on Circuits and Systems-I- IEEE Transactions on Computers- ACM Transactions on Reconfigurable Technology and Systems- IEEE Transactions on Emerging Topics in Computational Intelligence

Edge Computing and Inclusive Innovation: Shaping the Future of Smart and Safe Technologies

In an increasingly connected world, the potential for technology to drive societal progress is vast. However, to fully unlock this potential, we must address two crucial areas: the optimization of computational frameworks and the promotion of inclusivity in emerging technological environments. This PhD proposal focuses on these two pillars: the development of edge computing platforms and the application of inclusivity within the context of safe and smart cities.

The first pillar centers on edge computing platforms, with a focus not only on smart cities but also on broader applications. Edge computing, which allows data to be processed closer to its source, offers significant advantages across a wide range of fields, from healthcare and industrial automation to autonomous vehicles and IoT ecosystems. By reducing latency and improving response times, edge computing is essential for real-time decision-making and enhancing the performance of critical systems. This research will investigate how to optimize edge computing platforms for scalable, secure, and efficient use in multiple sectors, exploring its impact on data-intensive applications, distributed networks, and decentralized infrastructure. Key areas of study will include resource management, security, and the development of algorithms that maximize edge computing's potential across various industries.

The second pillar emphasizes the importance of inclusivity, particularly in the context of smart and safe cities. As cities evolve to incorporate smarter technologies, it is critical to ensure that all residents, including those with disabilities or from marginalized communities, can benefit from these advancements. This research will explore how edge computing can be used to develop inclusive applications, such as adaptive urban infrastructure, accessible communication systems, and personalized public safety services. By focusing on inclusive design, this pillar aims to create smart city solutions that are equitable and accessible to all, ensuring that technology contributes to the well-being of diverse populations.

By integrating these two pillars, this PhD proposal seeks to create a holistic approach where the power of edge computing not only enhances the performance of diverse applications but also fosters inclusivity in the development of future technologies. The outcome will be a framework that applies edge computing for efficient and accessible solutions across sectors, with a particular focus on making smart cities safer and more inclusive for everyone.

Year 1: Foundational Research and Framework Development
The first year will focus on establishing the theoretical and practical foundation for the research. A thorough literature review will analyze the state of the art in edge computing platforms and inclusive technologies. Technical requirements for optimizing edge computing frameworks in smart city contexts will be identified, and a preliminary design for resource management and security protocols will be developed. Simultaneously, specific use cases relevant to inclusivity, such as adaptive urban infrastructure and accessible communication systems, will be defined. This foundational work will set the stage for developing prototypes in subsequent years.

Year 2: Prototype Implementation and Performance Evaluation
Building on the groundwork laid in the first year, the second year will involve the development and implementation of prototypes for edge computing solutions addressing the identified use cases. These prototypes will be evaluated in real-world environments to assess their performance and scalability. Collaborations with stakeholders, such as urban planners, industry experts, and community organizations, will be essential to refine these solutions and ensure they align with the needs of diverse populations. Performance metrics such as latency, efficiency, and user accessibility will be rigorously measured, providing critical insights for further optimization.

Year 3: Scaling Prototypes and Expanding Applications
The third year will focus on enhancing the scalability and robustness of the developed solutions, ensuring they can be applied across various sectors and larger-scale scenarios. Research activities from the second year will continue, with particular attention to integrating the solutions into more complex environments and extending their application to additional use cases beyond the initial pilot projects. Further iterations of prototypes will be deployed, incorporating feedback from real-world testing and stakeholder collaboration. This iterative process will ensure that the solutions are adaptable, inclusive, and ready for broader adoption. The findings will also be disseminated through publications in high-impact journals and presentations at leading conferences.

The research projects correlated to the present proposals in the last years are:
[2024] PoC Instrument ?Safe Smart City: Rilevare Violenza e Richieste di Aiuto in Tempo Reale Tramite i Dispositivi di Videosorveglianza ? Fondazione Compagnia di San Paolo
[2023] HONEY ? Hybrid ONline tEchnologY fr particle therapy ? PRIN 2023

The research program will target the following conference venues: 
- IEEE International Conference on Edge Computing 
- IEEE International Conference on Smart Computing
- IEEE International Conference on Smart Systems and Edge Computing
- EEE Symposium on Edge Intelligence
- ACM/IEEE Symposium on Edge Computing

The research program will target the following journals: 
- IEEE Transactions on Emerging Topics in Computational Intelligence
- Intelligent Systems with Applications
- IEEE Information Systems