Enhancing Privacy Protection for Federated Learning with Distributed Differential Privacy

The primary objective of this study is to introduce an innovative framework for privacy-preserving federated learning (FL) by utilizing distributed differential privacy through secure aggregation (DDP-SA) to tackle the issue of privacy breaches during the FL training phase. Recent studies have demonstrated that traditional FL methods do not offer adequate privacy assurances, as malicious entities can deduce sensitive information about the training data from local model parameters or gradients, known as privacy inference attacks. The core concept behind DDP-SA is to integrate local differential privacy (LDP) and secure multi-party computation (MPC) to safeguard the gradients of local clients throughout the FL training process. Comprehensive experiments have been conducted to demonstrate that the proposed DDP-SA approach can deliver enhanced privacy safeguards in the FL process in comparison to either a solely LDP or solely MPC approach, while still maintaining a reasonable level of efficiency and accuracy.