Towards Secure Federated Learning: Enhancing Privacy and Robustness in Decentralized AI Systems
Rony Louis, Nader Bakir, Khouloud Samrouth, Vadim Pak · 2024
The traditional centralized training of artificial intelligence (AI) models faces growing privacy concerns as data increasingly resides in isolated silos and societal awareness of data privacy rises. Besides developing accurate global models, ensuring privacy in Machine Learning (ML) systems is now essential. Federated Learning (FL) has emerged as a powerful paradigm, enabling collaborative model training across decentralized devices, crucial in scenarios where sharing personal user data is risky. However, this decentralized approach introduces new privacy risks, particularly through inference attacks where adversaries can intercept and analyze models to extract sensitive information. Hence, there is still a need to protect FL systems. In this paper, we propose to encrypt the model parameters (weights and biais) updates exchanged between the FL server and clients. By implementing cryptographic protocols, we demonstrate how encryption can secure model parameters and predictions, preventing unauthorized access while maintaining model integrity. Our results indicate that incorporating encryption into FL offers a reasonable trade-off between security and performance, enhancing data privacy and overall security.