Scholarlink

This paper proposes FedCLUS, a privacy-preserving FL framework with dynamic client clustering to address data heterogeneity and data privacy challenges in medical domains. By leveraging pseudo-data generated through a Conditional WGAN-GP, the server estimates model similarity and clusters clients accordingly. This enables adaptive and personalized model aggregation without sharing sensitive patient data. The approach improves clustering accuracy, model robustness and performance under non-IID conditions. Experimental results on a sleep posture classification task demonstrate that FedCLUS outperforms existing FL methods in both accuracy and privacy protection across varying privacy budgets and gradient clipping settings, offering a scalable and secure solution for healthcare applications.

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 Privacy-Preserving Federated Learning Framework with Adaptive Client Clustering for Healthcare AI

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yexuan Hu,
  • Yan Sun,
  • Shunan Wu,
  • Qiliang Li,
  • Zhaoming Li

摘要

This paper proposes FedCLUS, a privacy-preserving FL framework with dynamic client clustering to address data heterogeneity and data privacy challenges in medical domains. By leveraging pseudo-data generated through a Conditional WGAN-GP, the server estimates model similarity and clusters clients accordingly. This enables adaptive and personalized model aggregation without sharing sensitive patient data. The approach improves clustering accuracy, model robustness and performance under non-IID conditions. Experimental results on a sleep posture classification task demonstrate that FedCLUS outperforms existing FL methods in both accuracy and privacy protection across varying privacy budgets and gradient clipping settings, offering a scalable and secure solution for healthcare applications.