Scholarlink

<p>Personalized Federated Learning (pFL) has emerged as a promising approach for adapting global models to heterogeneous client data distributions. However, existing pFL methods often suffer from catastrophic forgetting and unstable adaptation due to skewed or disjoint client distributions. To address these issues, we propose FedALA-WR, a novel framework that integrates adaptive attention-based aggregation with a weighted memory replay mechanism, allowing clients to reinforce valuable past knowledge during local training. A weighted loss function balances contributions from both the memory buffer and current local data. Additionally, an uncertainty-aware update strategy for memory replay enhances the sample efficiency and training stability. Despite the computational overhead caused by memory replay, FedALA-WR maintains competitive efficiency due to faster convergence, benefiting from an auto-stopping mechanism based on model stability. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that FedALA-WR consistently outperforms baseline FL and state-of-the-art pFL methods in accuracy and convergence speed. These results highlight its effectiveness in handling client heterogeneity without compromising scalability.</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

FedALA-WR: personalized federated learning with adaptive local aggregation and weighted replay

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Tinku Singh,
  • Nguyen Khoa,
  • Bhawnesh Kumar,
  • Taehong Kim

摘要

Personalized Federated Learning (pFL) has emerged as a promising approach for adapting global models to heterogeneous client data distributions. However, existing pFL methods often suffer from catastrophic forgetting and unstable adaptation due to skewed or disjoint client distributions. To address these issues, we propose FedALA-WR, a novel framework that integrates adaptive attention-based aggregation with a weighted memory replay mechanism, allowing clients to reinforce valuable past knowledge during local training. A weighted loss function balances contributions from both the memory buffer and current local data. Additionally, an uncertainty-aware update strategy for memory replay enhances the sample efficiency and training stability. Despite the computational overhead caused by memory replay, FedALA-WR maintains competitive efficiency due to faster convergence, benefiting from an auto-stopping mechanism based on model stability. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that FedALA-WR consistently outperforms baseline FL and state-of-the-art pFL methods in accuracy and convergence speed. These results highlight its effectiveness in handling client heterogeneity without compromising scalability.