Scholarlink

<p>Federated learning (FL) enables privacy-preserving model training across dispersed devices. Meanwhile, it faces non-IID data and backdoor attacks, which still has issues with consistency, efficiency, and security. In this work, we propose HBIpFL, a hypernetwork-based personalization and poisoning-free federated learning scheme to improve security and personalization in FL. HBIpFL dramatically reduces communication overhead without sacrificing performance. To achieve this, we leverage a hypernetwork-based parameter classifier to dynamically analyze and only upload important model parameters. Furthermore, we adopt local gradient ascent mechanism to track training loss trends and identify possible backdoor intrusions, guaranteeing the resilience and dependability of the global model. We compare HBIpFL with existing state-of-the-art schemes in the context of accuracy, communication overhead, and security. Comprehensive analysis indicate that HBIpFL provides a secure and efficient FL framework for diverse data distributions and privacy concerns.</p>

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

HBIpFL: hypernetwork and backdoor isolation personalized federated learning

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Jiankang Chen,
  • Haipeng Jiang,
  • Yuxin Xi,
  • Chunxiao Li,
  • Enliang Xu,
  • Shiyuan Xu,
  • Jing Wang

摘要

Federated learning (FL) enables privacy-preserving model training across dispersed devices. Meanwhile, it faces non-IID data and backdoor attacks, which still has issues with consistency, efficiency, and security. In this work, we propose HBIpFL, a hypernetwork-based personalization and poisoning-free federated learning scheme to improve security and personalization in FL. HBIpFL dramatically reduces communication overhead without sacrificing performance. To achieve this, we leverage a hypernetwork-based parameter classifier to dynamically analyze and only upload important model parameters. Furthermore, we adopt local gradient ascent mechanism to track training loss trends and identify possible backdoor intrusions, guaranteeing the resilience and dependability of the global model. We compare HBIpFL with existing state-of-the-art schemes in the context of accuracy, communication overhead, and security. Comprehensive analysis indicate that HBIpFL provides a secure and efficient FL framework for diverse data distributions and privacy concerns.