Scholarlink

<p>Federated Learning (FL) is a privacy-preserving distributed learning framework widely used in cross-device data modeling. However, traditional FL methods often suffer from high communication costs and performance degradation due to feature heterogeneity. To address these issues, we propose a federated adaptive prototype knowledge distillation method (FedPKD), which abstracts local models into prototype representations and soft labels for server-side aggregation, significantly reducing communication overhead. A similarity-aware adaptive aggregation strategy is introduced to generate personalized global knowledge for each client, which is distilled back to enhance local adaptation. Experiments on MNIST, FEMNIST, and CIFAR-10 show that FedPKD achieves superior accuracy, efficiency, and robustness under heterogeneous settings.</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

Fedpkd: a federated adaptive prototype knowledge distillation method

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Yongli Yang,
  • Shujie Ge,
  • Shiqiang Zhang,
  • Yang Cao,
  • Xiaonan Chen,
  • Ji Li

摘要

Federated Learning (FL) is a privacy-preserving distributed learning framework widely used in cross-device data modeling. However, traditional FL methods often suffer from high communication costs and performance degradation due to feature heterogeneity. To address these issues, we propose a federated adaptive prototype knowledge distillation method (FedPKD), which abstracts local models into prototype representations and soft labels for server-side aggregation, significantly reducing communication overhead. A similarity-aware adaptive aggregation strategy is introduced to generate personalized global knowledge for each client, which is distilled back to enhance local adaptation. Experiments on MNIST, FEMNIST, and CIFAR-10 show that FedPKD achieves superior accuracy, efficiency, and robustness under heterogeneous settings.