Scholarlink

<p>Recently, Heterogeneous Federated Learning (HtFL) has become a key paradigm for handling diverse model architectures and non-IID data across clients. Prototype-based methods, which share lightweight class prototypes to reduce communication costs, have attracted increasing attention. However, existing approaches suffer from “prototype contamination”—local prototypes uploaded by clients often carry client-specific biases, leading to suboptimal global aggregation and degraded local performance. To address this issue, we propose FedUPL, a novel HtFL framework that integrates a lightweight feature decoupling model (DM) and dual-classifier heads (DCH). The DM dynamically generates personalized and generalizable vectors to decompose extracted features into client-specific and globally shareable components, enabling unbiased prototype construction for global sharing. The DCH further separates local task optimization from global alignment through dedicated pathways without mutual interference, while a contrastive loss guides the decoupling process through global–local prototype constraints. Experiments on four benchmark datasets (Cifar10, Cifar100, Flowers102, and Tiny-ImageNet) demonstrate that FedUPL outperforms nine state-of-the-art methods, with the highest accuracy improvement reaching 4.89%. It also maintains a 5.2% advantage in fully heterogeneous scenarios while preserving the inherent communication efficiency of prototype-based approaches.</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

FedUPL: unbiased prototype learning via lightweight feature decoupling and dual-classifier heads in heterogeneous federated learning

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Xiaojiang Cui,
  • Yingxun Fu,
  • Li Ma

摘要

Recently, Heterogeneous Federated Learning (HtFL) has become a key paradigm for handling diverse model architectures and non-IID data across clients. Prototype-based methods, which share lightweight class prototypes to reduce communication costs, have attracted increasing attention. However, existing approaches suffer from “prototype contamination”—local prototypes uploaded by clients often carry client-specific biases, leading to suboptimal global aggregation and degraded local performance. To address this issue, we propose FedUPL, a novel HtFL framework that integrates a lightweight feature decoupling model (DM) and dual-classifier heads (DCH). The DM dynamically generates personalized and generalizable vectors to decompose extracted features into client-specific and globally shareable components, enabling unbiased prototype construction for global sharing. The DCH further separates local task optimization from global alignment through dedicated pathways without mutual interference, while a contrastive loss guides the decoupling process through global–local prototype constraints. Experiments on four benchmark datasets (Cifar10, Cifar100, Flowers102, and Tiny-ImageNet) demonstrate that FedUPL outperforms nine state-of-the-art methods, with the highest accuracy improvement reaching 4.89%. It also maintains a 5.2% advantage in fully heterogeneous scenarios while preserving the inherent communication efficiency of prototype-based approaches.