Scholarlink

<p>As a distributed machine learning paradigm, federated learning (FL) involves data samples that are distributed across multiple clients without sharing their local data. In a typical FL setting, the central server randomly selects clients to participate in training. However, studies have shown that random selection can lead to inefficiencies and reduced accuracy due to statistical heterogeneity. Therefore, in partially participated FL, it is crucial to reasonably select a subset of clients whose updates approximately represent the global data distribution. A novel gradient diversity metric is proposed in this paper to select clients by maximizing the gradient diversity of the client subset. To ensure fairness and prevent performance degradation due to gradient conflicts, a method is also proposed to eliminate gradient conflicts by modifying the gradient direction and magnitude. Through the joint optimization of client selection and gradient correction, this paper proposes a new FL framework called FedDC. Validation in heterogeneous environments demonstrates that FedDC improves training efficiency and model accuracy. Experimental results show that FedDC stabilizes global model updates, promotes uniform performance among clients, achieves a 2.3x training speedup, and improves accuracy by 17.3%.</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

FedDC: a federated learning framework to enhance gradient diversity and mitigate conflicts

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Yazhi Liu,
  • Haonan Xia,
  • Wei Li

摘要

As a distributed machine learning paradigm, federated learning (FL) involves data samples that are distributed across multiple clients without sharing their local data. In a typical FL setting, the central server randomly selects clients to participate in training. However, studies have shown that random selection can lead to inefficiencies and reduced accuracy due to statistical heterogeneity. Therefore, in partially participated FL, it is crucial to reasonably select a subset of clients whose updates approximately represent the global data distribution. A novel gradient diversity metric is proposed in this paper to select clients by maximizing the gradient diversity of the client subset. To ensure fairness and prevent performance degradation due to gradient conflicts, a method is also proposed to eliminate gradient conflicts by modifying the gradient direction and magnitude. Through the joint optimization of client selection and gradient correction, this paper proposes a new FL framework called FedDC. Validation in heterogeneous environments demonstrates that FedDC improves training efficiency and model accuracy. Experimental results show that FedDC stabilizes global model updates, promotes uniform performance among clients, achieves a 2.3x training speedup, and improves accuracy by 17.3%.