Scholarlink

Federated Learning (FL) has gained widespread adoption due to its privacy-preserving capabilities and collaborative learning framework. Nevertheless, in edge computing deployments, significant statistical heterogeneity across client data distributions combined with stringent communication constraints severely undermines the effectiveness of cross-client collaborative model training. To solve this problem, we introduce a unified framework for personalized federated learning that integrates partial network update with fine-grained parameter-level adaptation. We update network layers sequentially across communication rounds, alleviating layer mismatch and adopt a dynamic freezing strategy to enhance convergence. Concurrently, we employ a Fisher information-based sensitivity metric to identify sensitive parameters for each client. These parameters are selectively aggregated among clients with similar data distributions, while the remaining parameters are shared globally. Theoretical analysis reveals that this hybrid approach mitigates client heterogeneity at multiple scales. Our algorithmic contributions emphasize personalization with limited communication cost. Experimental results show that our method has better performance compared to state-of-the-art methods.

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

FedMini: A Federated Learning Framework with Partial Network Updates and Parameter-Wise Personalization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zilong Guo,
  • Jie Zhang,
  • Canran Li,
  • Huiyong Liu,
  • Lanlan Rui

摘要

Federated Learning (FL) has gained widespread adoption due to its privacy-preserving capabilities and collaborative learning framework. Nevertheless, in edge computing deployments, significant statistical heterogeneity across client data distributions combined with stringent communication constraints severely undermines the effectiveness of cross-client collaborative model training. To solve this problem, we introduce a unified framework for personalized federated learning that integrates partial network update with fine-grained parameter-level adaptation. We update network layers sequentially across communication rounds, alleviating layer mismatch and adopt a dynamic freezing strategy to enhance convergence. Concurrently, we employ a Fisher information-based sensitivity metric to identify sensitive parameters for each client. These parameters are selectively aggregated among clients with similar data distributions, while the remaining parameters are shared globally. Theoretical analysis reveals that this hybrid approach mitigates client heterogeneity at multiple scales. Our algorithmic contributions emphasize personalization with limited communication cost. Experimental results show that our method has better performance compared to state-of-the-art methods.