Differentially Private Federated Learning with Heterogeneous Group Privacy
Mingna Jiang, Linna Wei, Guoyue Cai, Xuangou Wu · 2023
Federated Learning (FL) can collaboratively learn Machine Learning (ML) models without sharing private data, effectively protecting data privacy. Empirical studies have shown that model parameters can also lead to privacy leakage. Differential Privacy (DP) has become a promising privacy-preserving solution to ensuring client privacy by introducing noise into the model parameters. However, a trade-off exists between the degree of privacy protection and model performance in DP-enhanced FL. Existing works assign the same level of privacy protection to all clients, ignoring the privacy requirements of clients. To address this challenge, we propose FedHGP, a differentially private FL approach with heterogeneous group privacy. Our approach divides clients into groups and applies Heterogeneous Differential Privacy (HDP) to each group individually. First, FedHGP groups clients into different levels of privacy protection based on the similarity of their data distributions. Then, various groups are added at varying noise levels and perform model aggregation separately. Finally, we conduct experiments on three benchmark datasets. The results show that our approach can effectively improve model performance when clients’ data and privacy needs are heterogeneous.