Secure Federated Learning Schemes Based on Multi-Key Homomorphic Encryption
Wenxiu Ding, Hongjiang Guo, Zheng Yan, Mingjun Wang · 2024
Federated learning (FL) has effectively solved the dilemma of "data silos" and has been widely applied into various fields. However, in real-life, user data is often Not Independent and Identically Distributed (Non-IID), and eavesdroppers can still infer users’ private data from various parameters transmitted in FL. Nevertheless, few privacy-preserving FL schemes focus on data heterogeneity, and offline problem causes more accuracy loss under Non-IID. This paper introduces a multi-key homomorphic encryption algorithm to construct two privacy-preserving FL schemes based on FedProx and SCAFFOLD, namely EMKProx and EMKSCAF. The two schemes both incorporate an anti-offline mechanism and take diverse measures to mitigate Non-IID’s impact on convergence speed, where EMKProx inserts a proximal term into loss function which is more suitable for poor hardware condition and EMKSCAF adds control variate to meet higher model performance requirement. We give security analyses to prove their security. In addition, we conduct simulations and comparisons with FL schemes using other privacy protection mechanisms, which show that our schemes behave more efficient and gain higher accuracy than the others.