Scholarlink

Although traditional federated learning (FL) methods excel in addressing label skew issues through local model updates and personalized learning strategies, they fall short in handling feature skew, which limits the model’s generalization capability across different domains. We propose FedPA, an innovative framework designed to significantly enhance model generalization performance in feature-skewed environments by utilizing prototypes to extract domain-invariant features. FedPA employs advanced prototype aggregation strategies and a feature adapter network trained end-to-end on the server side, ensuring the extraction of consistent, unbiased prototypes. On the client side, it enhances model accuracy and stability through alignment with these unbiased prototypes. Experimental evaluations confirm that FedPA significantly surpasses existing FL methods in terms of accuracy and robustness across multiple datasets. Notably, on the Digits-5 dataset, FedPA achieves a 5% improvement in accuracy compared to the standard FedAvg method and other advanced techniques. Furthermore, on the Office-Caltech-10 dataset, the global model’s accuracy is enhanced by 11.9%. These findings underscore FedPA’s effectiveness as a solution for cross-domain applications in FL.

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

FedPA: Unbiased Prototype Alignment in Federated Learning to Mitigate Feature Skew

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Qiuli Chen,
  • Xin Chen,
  • Feng Li,
  • Xiangli Yang

摘要

Although traditional federated learning (FL) methods excel in addressing label skew issues through local model updates and personalized learning strategies, they fall short in handling feature skew, which limits the model’s generalization capability across different domains. We propose FedPA, an innovative framework designed to significantly enhance model generalization performance in feature-skewed environments by utilizing prototypes to extract domain-invariant features. FedPA employs advanced prototype aggregation strategies and a feature adapter network trained end-to-end on the server side, ensuring the extraction of consistent, unbiased prototypes. On the client side, it enhances model accuracy and stability through alignment with these unbiased prototypes. Experimental evaluations confirm that FedPA significantly surpasses existing FL methods in terms of accuracy and robustness across multiple datasets. Notably, on the Digits-5 dataset, FedPA achieves a 5% improvement in accuracy compared to the standard FedAvg method and other advanced techniques. Furthermore, on the Office-Caltech-10 dataset, the global model’s accuracy is enhanced by 11.9%. These findings underscore FedPA’s effectiveness as a solution for cross-domain applications in FL.