后门
计算机科学
稳健性(进化)
离群值
计算机安全
钥匙(锁)
鉴定(生物学)
直觉
联合学习
数据挖掘
人工智能
机器学习
实证研究
数据建模
软件错误
作者
Pei Ye,Yuqing Li,Kun He,Qiao Li,Tianjie Qin,Xiong Wang,Kaige Yang,Chujun Zhang,Jing Chen
标识
DOI:10.1109/tifs.2025.3643155
摘要
Existing backdoor defense methods for federated learning (FL) usually try to distinguish between benign and malicious clients. The key insight is that benign clients are densely distributed, whereas malicious clients tend to be outliers outside this distribution. However, this only holds when data is independent and identically distributed (IID), and the effectiveness of these methods under non-IID data has not been systematically examined. In this paper, we present a comprehensive systematization of FL backdoor defense by breaking down its overall pipeline into three key components, i.e., metrics for evaluating clients, techniques for amplifying the difference between benign and malicious clients, and mechanisms for identifying malicious clients. We conduct an empirical study of FL backdoor defense methods under non-IID data settings to explore whether benign and malicious clients can be fully distinguished. Experimental results show that the defense performance degrades significantly when data is non-IID. Our results also reveal how evaluation metrics, amplification techniques and identification mechanisms perform under diverse settings. Contrary to the established belief, we further conclude that these defenses have inherent shortcomings, due to lack of stability and robustness in detecting malicious clients. We believe that our findings can better facilitate the development of FL backdoor defenses.
科研通智能强力驱动
Strongly Powered by AbleSci AI