FEDERATED LEARNING FRAMEWORK BASED ON TRIMMED MEAN AGGREGATION RULES

Wang Tian Xiang, Meiyue Shao, Yanwei Fu, Riheng Jia, Feilong Lin, Zhonglong Zheng

Keywords: Tmean，Byzantine attack，Byzantine-resilient

Abstract: This paper studies the problem of information security in the distributed learning framework. In particular, we consider the clients will always be attacked by Byzantine nodes and poisoning in the federated learning. Typically, aggregation rules are utilized to protect the model from the attacks in federated learning. The classical aggregation methods are Krum(·) and Mean(·), which however, are not capable enough to deal with Byzantine attacks in which general deviations and multiple clients are attacked at the same time. We propose new aggregation rules, Tmean(·), to the federated learning algorithm, and propose a federated learning framework based on Byzantine-resilient aggregation algorithm. Our novel Tmean(·) rules are derived from Mean(·) by appropriately trimming some of the values before averaging them. Theoretically, we provide rigorous theoretical proof and understanding of Tmean(·). Extensive experiments validate the effectiveness of our approaches.

One-sentence Summary: We analyze the Tmean(.) aggregation rule and use experiments to prove its robustness.