Byzantine-robust federated learning over Non-IID data

Abstract

The malicious attacks of Byzantine nodes in federated learning was studied over the non-independent and identically distributed dataset , and a privacy protection robust gradient aggregation algorithm was proposed.A reference gradient was designed to identify “poor quality” shared gradients in model training, and the influence of heterogeneity data on Byzantine node recognition was reduced by reputation evaluation.Meanwhile, the combination of homomorphic encryption and random noise obfuscation technology was introduced to protect user privacy in the process of model training and Byzantine node recognition.Finally, through the evaluation over the real-world datasets, the simulation results show that the proposed algorithm can accurately and efficiently identify Byzantine attack nodes while protecting user privacy and has good convergence and robustness.

Keywords

• federated learning;Byzantine attack;Non-IID;privacy protection;homomorphic encryption