Scholarlink

<p>Federated Learning (FL) significantly advances the field of distributed machine learning by allowing devices to collaboratively train models without the need to exchange sensitive local data, thus maintaining privacy and security. However, FL faces serious challenges in heterogeneous network environments where devices vary greatly in system resources (system heterogeneity) and data distributions (data heterogeneity), which impact the efficiency and effectiveness of the model training. To address these issues, we present FedCSGA, a novel FL client selection algorithm based on the genetic algorithm with adaptive genetic operators to maximize the number of selected clients within a training time deadline per round. Furthermore, we introduce the model accuracy into the fitness function of FedCSGA and enable the algorithm to handle the system and data heterogeneity simultaneously. Our comprehensive experiments demonstrate that the proposed FedCSGA algorithm can significantly outperform several state-of-the-art baselines and increase the number of clients selected per round by 45.6%, 54.3%, the model accuracy by 2.4%, 7.7% on average in both IID and non-IID settings, respectively.</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

Optimizing Client Selection in Federated Learning Base on Genetic Algorithm

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Jiagao Wu,
  • Hongyan Ji,
  • Jing Yi,
  • Linfeng Liu

摘要

Federated Learning (FL) significantly advances the field of distributed machine learning by allowing devices to collaboratively train models without the need to exchange sensitive local data, thus maintaining privacy and security. However, FL faces serious challenges in heterogeneous network environments where devices vary greatly in system resources (system heterogeneity) and data distributions (data heterogeneity), which impact the efficiency and effectiveness of the model training. To address these issues, we present FedCSGA, a novel FL client selection algorithm based on the genetic algorithm with adaptive genetic operators to maximize the number of selected clients within a training time deadline per round. Furthermore, we introduce the model accuracy into the fitness function of FedCSGA and enable the algorithm to handle the system and data heterogeneity simultaneously. Our comprehensive experiments demonstrate that the proposed FedCSGA algorithm can significantly outperform several state-of-the-art baselines and increase the number of clients selected per round by 45.6%, 54.3%, the model accuracy by 2.4%, 7.7% on average in both IID and non-IID settings, respectively.