MTP: A Minimal Task Perturbation Framework for Mitigating Poisoning Attacks in Federated Learning

Federated Learning achieves privacy protection by allowing different participants to jointly train models. However, this characteristic also makes Federated Learning susceptible to data poisoning attacks. Many existing defense methods can successfully defend against various data poisoning attacks, but they are often only effective against one of the independent and identically distributed (IID) or non-independent and identically distributed (non-IID) scenarios. In this paper, we propose a novel minimal task perturbation framework named MTP for mitigating data poisoning attacks in both IID and non-IID scenarios. To efficiently identify potential attackers, MTP utilizes a weight similarity ensemble mechanism based on four distance algorithms and clusters clients according to the fused weight similarity. Moreover, MTP uses a clipped weight aggregation to minimize the impact on the performance of main task. Experimental results show that the proposed MTP outperforms state-of-the-art methods in stability and security, and maintain higher accuracy of the main task.