Distributed Differential Privacy for Federated Learning: A Privacy-Enhancing Approach
Hao Zhou, Jie Kong · 2024
To enhance data protection during federated learning training, this paper proposes a federated learning approach based on distributed differential privacy. In the distributed framework, a secure shuffle model is adopted as a trusted aggregator, and the shuffler randomly mixes reports and forwards them to the server for final analysis. In addition, the concept of Renyi divergence is introduced to more precisely control privacy loss. The results show that this method not only provides effective privacy protection for the federated learning training process, but also maintains good model performance, especially on the MINIST dataset.