Block-BRFL: Blockchain-based High Byzantine Robust Federated Learning Method
Wenjie Xiao, Huazhen Zhong, Yangchen Dong, Xuehai Tang, Xiaodan Zhang, Jizhong Han · 2024
Recently, there has been significant attention on the security research of federated learning, particularly in leveraging traditional Byzantine fault tolerance techniques to address Byzantine attacks. However, many studies often make the assumption of a trusted central server. In reality, the central server can be malicious, posing a threat to the security of Byzantine fault tolerance techniques. This concern becomes even more serious when dealing with non-IID data. In this paper, we propose a decentralized and secure federated learning method called Block-BRFL, which aims to address the issue of an untrusted central server and effectively combat Byzantine attacks. Block-BRFL achieves this by incorporating blockchain technology to facilitate a secure and coordinated machine learning process among participating clients. Additionally, Block-BRFL employs committee aggregation algorithms based on feedback mechanisms to mitigate the impact of malicious attacks, both at the local model level and the global model level, thereby enhancing the defense against Byzantine attacks. Furthermore, we evaluate the performance of Block- BRFL on real-world datasets, and the results demonstrate a significant improvement in accuracy compared to the original federated average algorithm. Specifically, under the presence of 20% Byzantine attack clients, Block-BRFL achieves a remarkable increase in accuracy by approximately 32%.