Scholarlink

<p>As a decentralized machine learning paradigm, federated learning facilitates collaborative model training while inherently preserves client data privacy. However, in the non-IID settings, conventional federated learning approaches encounter poor model generalization. Group-based aggregation can alleviate the issues caused by non-IID data, but the lack of communication among groups leads to low data utilization. To address these challenges, a federated learning framework based on self-excluding aggregation (FedSEA) is proposed by integrating dynamic sparsity-driven client grouping, self-excluding aggregation and cross-group knowledge distillation. The FedSEA consists of three stages. First, privacy-preserving client grouping is achieved through sparsity analysis of feature mapping. Second, intra-group aggregation and self-exclusion aggregation are performed on grouped clients. Finally, cross-group knowledge distillation is conducted using knowledge distillation. Extensive experiments demonstrate that the FedSEA achieves superior performance on the Office-Home and CIFAR-10 datasets, with about 2% improvement in classification accuracy over the state-of-the-art 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

Personalized federated learning based on self-excluding aggregation with knowledge distillation

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Wen-Zheng Bi,
  • Wen-Jun Yu,
  • Li-Hua Gong

摘要

As a decentralized machine learning paradigm, federated learning facilitates collaborative model training while inherently preserves client data privacy. However, in the non-IID settings, conventional federated learning approaches encounter poor model generalization. Group-based aggregation can alleviate the issues caused by non-IID data, but the lack of communication among groups leads to low data utilization. To address these challenges, a federated learning framework based on self-excluding aggregation (FedSEA) is proposed by integrating dynamic sparsity-driven client grouping, self-excluding aggregation and cross-group knowledge distillation. The FedSEA consists of three stages. First, privacy-preserving client grouping is achieved through sparsity analysis of feature mapping. Second, intra-group aggregation and self-exclusion aggregation are performed on grouped clients. Finally, cross-group knowledge distillation is conducted using knowledge distillation. Extensive experiments demonstrate that the FedSEA achieves superior performance on the Office-Home and CIFAR-10 datasets, with about 2% improvement in classification accuracy over the state-of-the-art methods.