Scholarlink

<p>The proliferation of federated learning (FL), driven by the integration of data silos, enables collaborative model training without sharing confidential data. However, the Non-IID data distribution and uncertain contributions from different participants, particularly free-riders with low-quality datasets, tend to complicate federated aggregation. In this paper, we propose <span>FedHetero</span>, a novel FL framework guided by automated data source ranking to mitigate the impact of Non-IID settings on model accuracy and adaptively adjust the weights in the aggregation. We use deep reinforcement learning (DRL) to identify the most suitable contributions of different participants and update the weights in the federated aggregation. <span>FedHetero</span> is applicable in both centralized and decentralized federated learning settings. Experiments show that <span>FedHetero</span> achieves better efficiency and accuracy.</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

FedHetero: improving federated learning on heterogeneous data distribution through deep reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Shuo Sun,
  • Jiecong Wang,
  • Jie Chen,
  • Xianghong Tang,
  • Yulong Fang,
  • Shuhai Wang,
  • Jianwu Zheng,
  • Yongjian Huang,
  • Jianfeng Wang,
  • Wei-Tek Tsai,
  • Lin Liu,
  • Xin Liu,
  • Mingsheng Liu

摘要

The proliferation of federated learning (FL), driven by the integration of data silos, enables collaborative model training without sharing confidential data. However, the Non-IID data distribution and uncertain contributions from different participants, particularly free-riders with low-quality datasets, tend to complicate federated aggregation. In this paper, we propose FedHetero, a novel FL framework guided by automated data source ranking to mitigate the impact of Non-IID settings on model accuracy and adaptively adjust the weights in the aggregation. We use deep reinforcement learning (DRL) to identify the most suitable contributions of different participants and update the weights in the federated aggregation. FedHetero is applicable in both centralized and decentralized federated learning settings. Experiments show that FedHetero achieves better efficiency and accuracy.