Scholarlink

Current federated learning research assumes uniform model architecture across client devices. However, in edge computing, client resources often vary significantly. Using the same model may waste resources in powerful clients. Although adapting model complexity to client resources is reasonable, it introduces model architecture heterogeneity, which challenges many existing algorithms. Hence, we propose the FedTS algorithm, using data-free knowledge distillation to enable large models to assist smaller models, addressing both model and resource diversity in federated learning. The experiments show that our algorithm outperforms the baseline on multiple datasets(12.65% on CIFAR-10 and 8.73% on CINIC-10, compared to the baseline).

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

FedTS: Leveraging Teacher-Student Architecture in Federated Learning Against Model Heterogeneity in Edge Computing Scenarios

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zihong Lin,
  • Yucheng Tao,
  • Haopeng Chen

摘要

Current federated learning research assumes uniform model architecture across client devices. However, in edge computing, client resources often vary significantly. Using the same model may waste resources in powerful clients. Although adapting model complexity to client resources is reasonable, it introduces model architecture heterogeneity, which challenges many existing algorithms. Hence, we propose the FedTS algorithm, using data-free knowledge distillation to enable large models to assist smaller models, addressing both model and resource diversity in federated learning. The experiments show that our algorithm outperforms the baseline on multiple datasets(12.65% on CIFAR-10 and 8.73% on CINIC-10, compared to the baseline).