Anagrafe della ricerca

Cyber-Physical Systems penetrate different domains, from IoT to Automotive. Because CPSs are coupled with new and complex failure mechanisms due to nanometer-scale fabrication technology, a version consisting of many hardware and software components with mixed degrees of criticality, hardcoded redundancies in the design, and fixed runtime recovery policies are no longer sufficient. Moreover, they are subject to malicious attacks, requiring that the next-generation intelligent systems support new groundbreaking paradigms for a holistic approach towards learning-based resilience. This is particularly true in Automotive: when the system is not working correctly, it likely leads to potential hazards, compromising the safety of all the users. The primary purpose of this project is the pursuit of system dependability in the Automotive domain, complementing the well-known methodologies to design safe systems through a set of holistic data-driven learning-based approaches during the operation lifecycle of computing systems. The idea is to support a continuous learning approach, such that systems can continue to provide expected functionality despite malfunctioning components or cybersecurity attacks breakdown. This project aims at achieving the objective by exploring new architectures for continuous monitoring and learning. We aim at building a prototype based on a RISC-V microprocessor coupled with an Artificial Intelligence (AI) based hardware accelerator. The accelerator will include classification capabilities and reinforcement learning strategies to detect failures and cybersecurity threats, and apply proper countermeasures on the fly. Such architecture allows for building an intelligent system that exploits continuous monitoring for threat detection and To support the feasibility, we foresee a use case implementing an autonomous driving task deployed following all functional safety specifications by ISO 26262. The approach targets sensitive parts of the RISC-V microprocessor that might suffer from hardware faults, i.e., registers affected by stuck-at. The system will include an online real-time monitoring task, which will continuously observe RISC-V monitoring features, i.e., performance monitor counters (PMCs), to feed an Artificial Neural Network (ANN) model, which is trained for anomaly detection and running on a hardware accelerator for the sake of efficiency. The ANN enables the detection of safety-critical deviations of the system behavior either caused by HW faults or by cybersecurity attacks. Also, anomaly detection enables the activation of hardware or software, i.e., Operating system level, remedial actions. The latter are selected using a Deep Reinforcement Learning (DRL) based agent that determines the most appropriate sequence of actions to mitigate the fault and restore a typical behavior of the system without failing the functional safety of the system.