FedDiscrete: A Secure Federated Learning Algorithm Against Weight Poisoning

Yutong Dai, Xingjun Ma, Lichao Sun

Keywords: federated learning, weight poisoning defense

Abstract: Federated learning (FL) is a privacy-aware collaborative learning paradigm that allows multiple parties to jointly train a machine learning model without sharing their private data. However, recent studies have shown that FL is vulnerable to weight poisoning attacks. In this paper, we propose a probabilistic discretization mechanism on the client side, which transforms the client's model weight into a vector that can only have two different values but still guarantees that the server obtains an unbiased estimation of the client's model weight. We theoretically analyze the utility, robustness, and convergence of our proposed discretization mechanism and empirically verify its superior robustness against various weight-based attacks under the cross-device FL setting.