Scholarlink

<p>Federated learning (FL) is a distributed learning framework that enables multiple edge devices to collaboratively train models while preserving data privacy. However, in real-world applications, FL faces various challenges, such as data integrity issues, noisy data, and potential attacks, which can lead to erroneous model training and reduced performance. Previous methods primarily address label noise and attacks by adjusting weights or directly discarding clients that pose threats. However, these approaches often fail when the proportion of malicious clients is high or the noise level is severe. Moreover, simply removing clients contradicts the fundamental principle of FL, which encourages broad device participation. To address these challenges, we propose a novel framework, FedMBG, specifically designed to handle extreme noise and potential attacks in FL. FedMBG consists of two key stages: client selection and group optimization, which classifies and groups clients based on their updates, and inter-group personalization training and model optimization, which leverages information generated by malicious clients to guide other clients in enhancing model robustness, thereby mitigating the impact of noise and attacks. Experimental results demonstrate that FedMBG significantly improves model accuracy and robustness in extreme noise and high-attack environments, outperforming existing FL methods.</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

Multi-stage federated learning with group-wise bidirectional guidance

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Xiaohui Li,
  • Taicheng Bian,
  • Jin Yang,
  • Dehan Meng,
  • Yuhang Lu,
  • Xi Jiang,
  • Hongtao Liang

摘要

Federated learning (FL) is a distributed learning framework that enables multiple edge devices to collaboratively train models while preserving data privacy. However, in real-world applications, FL faces various challenges, such as data integrity issues, noisy data, and potential attacks, which can lead to erroneous model training and reduced performance. Previous methods primarily address label noise and attacks by adjusting weights or directly discarding clients that pose threats. However, these approaches often fail when the proportion of malicious clients is high or the noise level is severe. Moreover, simply removing clients contradicts the fundamental principle of FL, which encourages broad device participation. To address these challenges, we propose a novel framework, FedMBG, specifically designed to handle extreme noise and potential attacks in FL. FedMBG consists of two key stages: client selection and group optimization, which classifies and groups clients based on their updates, and inter-group personalization training and model optimization, which leverages information generated by malicious clients to guide other clients in enhancing model robustness, thereby mitigating the impact of noise and attacks. Experimental results demonstrate that FedMBG significantly improves model accuracy and robustness in extreme noise and high-attack environments, outperforming existing FL methods.