Scholarlink

In practical federated learning (FL) applications, client-side data is typically heterogeneous in distribution. Nevertheless, existing defense mechanisms exhibit limited robustness under such conditions and remain vulnerable to sophisticated poisoning attacks. To mitigate these compounded vulnerabilities, we propose a novel federated learning defense mechanism DPCFL, which leverages dual-personalized clustering to significantly improve robustness against both data heterogeneity and composite poisoning attacks. Initially, we devised an adaptive clustering-based grouping mechanism that segments clients into distributionally homogeneous clusters by extracting multi-dimensional statistical features from their model feature maps. Within each cluster, we subsequently implement personalized client-specific optimization and perform robust aggregation through dynamic weight allocation, substantially enhancing the robustness of federated learning under heterogeneous data conditions. Furthermore, we introduce a client anomaly detection metric based on the Tanimoto coefficient to improve the precision of the identification against composite attacks. Experimental results demonstrate that, compared to existing federated learning defense methods, our approach exhibits superior robustness under varying Non-IID conditions and composite poisoning attacks while significantly improving model accuracy performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Dual Personalized Clustering-Based Defense Against Composite Poisoning Attacks in Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • He Wang,
  • Zhen Xu,
  • Qian Tan,
  • Yan Zhang

摘要

In practical federated learning (FL) applications, client-side data is typically heterogeneous in distribution. Nevertheless, existing defense mechanisms exhibit limited robustness under such conditions and remain vulnerable to sophisticated poisoning attacks. To mitigate these compounded vulnerabilities, we propose a novel federated learning defense mechanism DPCFL, which leverages dual-personalized clustering to significantly improve robustness against both data heterogeneity and composite poisoning attacks. Initially, we devised an adaptive clustering-based grouping mechanism that segments clients into distributionally homogeneous clusters by extracting multi-dimensional statistical features from their model feature maps. Within each cluster, we subsequently implement personalized client-specific optimization and perform robust aggregation through dynamic weight allocation, substantially enhancing the robustness of federated learning under heterogeneous data conditions. Furthermore, we introduce a client anomaly detection metric based on the Tanimoto coefficient to improve the precision of the identification against composite attacks. Experimental results demonstrate that, compared to existing federated learning defense methods, our approach exhibits superior robustness under varying Non-IID conditions and composite poisoning attacks while significantly improving model accuracy performance.