Tamer Ahmed Eltaras

Dottorando in Ingegneria Informatica E Dei Sistemi , 38o ciclo (2022-2025)
Dipartimento di Automatica e Informatica (DAUIN)

Assegnista di Ricerca
Dipartimento di Automatica e Informatica (DAUIN)

Profilo

Dottorato di ricerca

Argomento di ricerca

Data Privacy and Security in Federated Learning: Attacks and Defense Mechanisms

Tutori

Presentazione della ricerca

Poster

Interessi di ricerca

Cybersecurity
Data science, Computer vision and AI

Biografia

In 2018, I earned my Master’s degree in Electronics from Politecnico di Torino, where my thesis, "A Flow Control Mechanism for Fully Adaptive Routing Algorithms in On-Chip Networks," explored innovative methods to optimize communication efficiency in complex network-on-chip systems. Currently, I am pursuing my PhD in Computer and Control Engineering at Politecnico di Torino as part of the SMILIES research group.

My research focuses on Data Privacy and Security in Federated Learning (FL), particularly the vulnerabilities and defenses within this privacy-preserving machine learning paradigm. Federated Learning enables decentralized model training across multiple clients without transferring raw data, addressing significant privacy concerns in data-driven applications. However, FL faces serious security risks, including gradient leakage attacks, where private data can be reconstructed from shared gradients, exposing sensitive information.

In my work, I explore how attackers might exploit gradients and initial model weights to reconstruct original private data and examine strategies to mitigate these threats. This includes developing R-CONV, a novel method that tackles data reconstruction from convolutional layers, and creating a secure, verifiable aggregation protocol for FL. These solutions are designed to counteract multiple types of attacks, striving to achieve a balance between privacy, security, and computational efficiency in federated systems. My research aims to provide robust protections to support the privacy needs of modern, distributed AI models, addressing both current and emerging challenges in federated learning security.

Ricerca

Gruppi di ricerca

Pubblicazioni

Pubblicazioni più recenti Vedi tutte le pubblicazioni su Porto@Iris