Privacy-Preserving Parameter Aggregation Scheme Based on Two Leaders for Federated Learning
Weiyu Wang, Zhuotao Lian, Tianhui Li, Kouichi Sakurai, Chunhua Su · 2024
Federated learning techniques have been developed to protect the privacy of personal data during multi-party collaboration. However, safeguarding the privacy of model updates and data transfers remains a challenge. Most existing methods rely on differential privacy and homomorphic encryption to protect personal data. However, the noise added by differential privacy can slow down the convergence and reduce the final accuracy of the model. Additionally, these methods often increase system complexity and computational burden, leading to high performance overhead and a challenging privacy budget. In this paper, we propose a new secure scheme for aggregating model parameters that does not require adding differential privacy noise or additional computational burden. We compare our scheme with differential privacy-based approach that effectively safeguards the model update and data transmission processes while preserving the model's original performance, ensuring high accuracy and a low loss rate. Furthermore, security proofs demonstrate that our scheme effectively protects client parameter privacy against eavesdropping attacks during transmission and maintains a low probability of successful collaborative attacks.