A Blockchain-based Grouped Federated Learning Scheme Against Malicious Clients

Hongle Guo, Yingchi Mao, Xiaoming He, Hua Nie

Abstract: The existence of malicious clients in federated learning leads to the decline of model performance or even training failure. Considering the huge advantages of the blockchain in the decentralization, feasibility and high reliability, many schemes based on blockchain have been proposed. However, due to the large number and scattered distribution of clients, it is difficult to backtrack malicious clients. In this paper, to reduce the time to backtrack the malicious clients and improve the accuracy of model training in the federated learning, we pro- pose a Blockchain-based Grouped Federated Learning Scheme (BGFLS). Specifically, to quickly and accurately backtrack the malicious clients, we present a grouping federated learning architecture, which applies blockchain technology to each group. Moreover, we design an algorithm to backtrack the malicious clients and propose a block structure to support backtracking. Theoretical analysis shows that the BGFLS can quickly backtrack malicious clients and protect shared parameters. We evaluate the BGFLS on MNIST and Fashion-MNIST datasets. The experimental results demonstrate that the accuracy of the BGFLS is higher than that of the GeoMed scheme around 0.08 and is higher than that of the Krum scheme around 0.07 when the proportion of the malicious clients is 20%. Meanwhile, the backtracking efficiency of the BGFLS is higher than that of the original blockchain.