Robust Federated Learning with Adaptable Learning Rate
Jingxuan Zhou, Zhengyi Zhong, Ji Wang, Xiongtao Zhang, Fuxu Chen, Chun Gang Yan · 2021
Federated learning(FL), one of the most popular distributed machine learning algorithms, allows participants to collaborate on updating models without sharing local data. At the same time, FL is vulnerable to malicious attacks due to the secrecy of the data. Backdoor attacks, as one of many malicious attacks, attempt to add a backdoor to the global model, which will lead to misclassification after activation. Backdoor attacks are difficult to detect. To defend against backdoor attacks, we designed a defense method deployed on the central server side. The method autonomously adjusts the learning rate based on the received parameter signs form all participants to reduce the impact of malicious updates. It is shown experimentally that the method can significantly reduce the impact of backdoor attacks while maintaining high model accuracy on clean datasets when the correct threshold is set.