Regulated Federated Learning against the Effects of Heterogeneity and Client Attacks

Federated Learning (FL), as a new privacy learning paradigm, can complete the learning task without compromising user privacy. With the enhancement of user data protection awareness, FL has received much attention and has been widely applied. However, the FL mechanism, where clients train models using personal data locally and exchange model updates instead of raw data, gives rise to new challenges. The problems caused by data heterogeneity and malicious client behaviors are universal in practical applications. We proposed a Regulated Federated Learning (RegulFL) that introduces a generator and uses weighted aggregations to regulate client model training and complete federated aggregation. Besides, we gave a weighted geometric median method to get client weights that can effectively reduce the influence of malicious clients in aggregation processes. Experimental results demonstrate that RegulFL is robust against malicious client attacks while effectively mitigating the effect of data heterogeneity.