FedCleanse: Cleanse the backdoor attacks in federated learning system

Siquan Huang, Yijiang Li, Chong Chen, Leyu Shi, Wentian Cai, Ying Gao

Abstract: Highlights•We propose FedCleanse, a novel defense deployed after aggregation, significantly superior to the previous DP-based or pruning-based method without extra “clean” data held in the server. Besides, FedCleanse can be combined with the defenses conducted before aggregation and achieve SOTA performance.•We propose a voting mechanism followed by a suppression process based on the “conductance value”, which eliminates the backdoor from the FL model while maintaining the benign performance.•We propose the perturbation process to mitigate the backdoors lurking in the benign neurons. This effectively prevents the backdoor from transferring to benign neurons during the FL training after suppression.•We perform an exhaustive evaluation of FedCleanse on various datasets, models, adversary settings, and attack methods. The results demonstrate new SOTA performance compared to the previous defenses. We further illustrate that our defense can plug and play with any pre-aggregation defense, achieving mutual benefits.

External IDs:dblp:journals/kbs/HuangLCSCG25