Scholarlink

Traditional federated learning (FL) methods like FedAvg often struggle with non-independent and identically distributed (non-IID) data, leading to poor performance and slow convergence. Clustered federated learning (CFL) addresses the non-IID data challenge in FL by grouping clients with similar data distributions into clusters and conducting FL within each cluster, thereby improving model accuracy and convergence speed. Existing CFL approaches face issues such as predefined cluster numbers, sensitivity to initial assignments, and high communication costs. This work proposes a robust and efficient CFL framework that adaptively identifies clusters based on client data similarity, mitigates outlier effects, and reduces communication overhead. Specifically, we introduce a unified framework for any FL topology and leverage self-expressive learning to understand client relationships, offering a simple yet effective CFL framework to address non-IID issues. Compared to advanced FL algorithms, our proposed method shows better performance, accelerated convergence, improved robustness against attacks and lower communication costs.

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

Robust and Efficient Clustered Federated Learning via Adaptive Self-Expressive Representations for Tackling Data Heterogeneity

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Wenlin Ou,
  • Tianchi Liao,
  • Chuan Chen

摘要

Traditional federated learning (FL) methods like FedAvg often struggle with non-independent and identically distributed (non-IID) data, leading to poor performance and slow convergence. Clustered federated learning (CFL) addresses the non-IID data challenge in FL by grouping clients with similar data distributions into clusters and conducting FL within each cluster, thereby improving model accuracy and convergence speed. Existing CFL approaches face issues such as predefined cluster numbers, sensitivity to initial assignments, and high communication costs. This work proposes a robust and efficient CFL framework that adaptively identifies clusters based on client data similarity, mitigates outlier effects, and reduces communication overhead. Specifically, we introduce a unified framework for any FL topology and leverage self-expressive learning to understand client relationships, offering a simple yet effective CFL framework to address non-IID issues. Compared to advanced FL algorithms, our proposed method shows better performance, accelerated convergence, improved robustness against attacks and lower communication costs.