Secure Decentralized Aggregation to Prevent Membership Privacy Leakage in Edge-Based Federated Learning
Meng Shen, Jing Wang, Jie Zhang, Qinglin Zhao, Bohan Peng, Tong Wu, Liehuang Zhu, Ke Xu
Abstract
Federated Learning (FL) is a machine learning approach that enables multiple users to share their local models for the aggregation of a global model, protecting data privacy by avoiding the sharing of raw data. However, frequent parameter sharing between users and the aggregator can incur high risk of membership privacy leakage. In this paper, we propose LiPFed, a computationally lightweight privacy preserving FL scheme using secure decentralized aggregation for edge networks. Under this scheme, we ensure privacy preservation on the aggregation side, and promote lightweight computation on the user side. By incorporating blockchain and additive secret sharing algorithm, we effectively protect the membership privacy of both local models and global models. Furthermore, the secure decentralized aggregation mechanism safeguards against potential compromises of the aggregator. Meanwhile, smart contract is introduced to identify malicious models uploaded by edge nodes and return trustworthy global models to users. Rigorous security analysis shows the effectiveness of this scheme in privacy preservation. Extensive experiments verify that LiPFed outperforms the state-of-the-art schemes in terms of training efficiency, model accuracy, and privacy preservation.