Exploration de l'apprentissage fédéré vérifiable et respectueux de la vie privée grâce à la confidentialité différentielle et aux protocoles cryptographiques

Abstract

Federated Learning (FL) has emerged as a distributed paradigm enabling multiple participants to collaboratively train machine learning models without sharing their raw data. By keeping data local, FL mitigates many privacy risks inherent to centralized learning architectures. However, despite this promise, recent research has revealed that exchanged gradients can still leak sensitive information about local datasets. Furthermore, most existing approaches rely on strong and often unrealistic trust assumptions toward the central server, while providing no means to verify whether privacy-preserving mechanisms have been correctly enforced. These limitations expose a critical gap between theoretical privacy guarantees and their practical implementation in real-world federated systems.This thesis investigates how to bridge this gap by combining differential privacy (DP) with cryptographic and verifiability protocols to achieve verifiable and trust-reduced federated learning. First, we explore the use of additive homomorphic encryption to protect client updates and minimize reliance on a trusted aggregator. Second, we introduce a non-interactive verifiability protocol based on zk-SNARKs and cryptographic hashes, allowing third parties to prove and verify the correct application of DP without revealing sensitive information. Finally, we propose ProoFed, a distributed framework that leverages secret sharing to decentralize noise generation and integrate verifiable aggregation proofs in zero knowledge, thereby eliminating single points of trust.