Unlearning Backdoor Attacks in Federated Learning
Chen Wu, Sencun Zhu, Prasenjit Mitra, Wei Wang
Abstract
Federated learning systems are constantly under the looming threat of backdoor attacks. Despite significant progress in mitigating such attacks, the challenge of effectively removing a potential attacker’s influence from the trained global model remains unresolved. In this paper, we present a novel federated unlearning method that is suitable for backdoor removal. By leveraging historical updates subtraction and knowledge distillation, our approach can maintain the models’s performance while completely removing the backdoors implanted by the attacker from the model. It can be seamlessly applied to various types of neural networks and does not require clients’ participation in the unlearning process. Through experiments on diverse computer vision and natural language processing datasets, we demonstrate the effectiveness and efficiency of our proposed method. The promising results obtained validate the potential of our approach to bolster the security of federated learning systems against backdoor threats.