Privacy-Preserving Asynchronous Federated Learning Framework in Distributed IoT
Xinru Yan, Yinbin Miao, Xinghua Li, Kim‐Kwang Raymond Choo, Xiangdong Meng, Robert H. Deng
Abstract
To solve the data island issue in the distributed Internet of Things (IoT) without privacy leakage, privacy-preserving federated learning (PPFL) has been extensively explored in both academic and industrial fields. However, existing PPFL solutions still suffer from a single point of failure and incur untrusted aggregation results caused by a malicious central server, and even cause a loss of model accuracy in an asynchronous setting. To solve these issues, we propose a privacy-preserving asynchronous federated learning scheme by using blockchain. Specifically, we use blockchain to address single points of failure and untrustworthy aggregation results, implement reliable model aggregation utilizing a practical byzantine fault-tolerant protocol in an asynchronous setting, and leverage differential privacy to improve system robustness. Formal security analysis and convergence analysis demonstrate that the proposed scheme is secure and robust, and extensive experiments demonstrate that our scheme can effectively ensure the accuracy of the system when compared with state-of-the-art schemes.