Reproduce “Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection” on OpenResearch

About this paper

Jiahao Xu, Zikai Zhang, Rui Hu

AlignIns, a new defense framework from the University of Nevada, Reno, detects backdoor attacks in Federated Learning by inspecting multi-granularity directional alignment of local model updates. It achieves significantly lower backdoor attack accuracy and higher robustness accuracy across various datasets, attack types, and non-IID settings compared to existing robust aggregation methods.

Key insights

• Malicious updates, by necessity of balancing main and backdoor tasks, exhibit subtle directional inconsistencies at both general and fine-grained (parameter sign) levels, which can be leveraged for detection.
• Focusing on the signs of 'important' parameters (those with largest absolute values) significantly improves the detection of stealthy attacks by reducing noise from less critical parameters.
• The combination of multi-granularity directional checks (TDA for general, MPSA for fine-grained) and robust statistical anomaly detection (MZ score) provides comprehensive protection that adapts to varying training stages and data heterogeneity.

Summary by alphaXiv.