Scholarlink

<p>The Non-IID characteristic of client data cause gradient vanishing or explosion in federated learning. It prevents the global model from converging. Client selection strategies have been proposed to address this problem. However, most existing methods are limited to a single selection criterion and fail to consider the impact of historical training information on client selection. In this paper, we propose an Adaptive Historical-Information-Based Client Selection (AHIBCS) algorithm. Firstly, we design an adaptive threshold adjustment mechanism. It dynamically computes and adjusts the threshold based on error change trends during training, optimizing the client selection strategy. Secondly, we propose a dual-factor selection strategy, which adapts to different training stages by using distinct selection criteria. Finally, we propose a historical information interaction mechanism. It allows client selection to consider both current performance metrics and historical performance trends, enabling flexible adjustments. This mechanism significantly enhances training stability and convergence efficiency. Experimental results show that, compared to the Pow-d method, AHIBCS improves the global model accuracy by 2.1%, 2.4%, and 2.7% on the COVID-19, FMNIST, and CIFAR-10 datasets, 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

AHIBCS: an adaptive historical-information-based client selection algorithm for Non-IID data

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Wei Liu,
  • Xu Wang,
  • Bin Wang,
  • Wentao Cui,
  • Zhao Tian,
  • Guangjun Zai

摘要

The Non-IID characteristic of client data cause gradient vanishing or explosion in federated learning. It prevents the global model from converging. Client selection strategies have been proposed to address this problem. However, most existing methods are limited to a single selection criterion and fail to consider the impact of historical training information on client selection. In this paper, we propose an Adaptive Historical-Information-Based Client Selection (AHIBCS) algorithm. Firstly, we design an adaptive threshold adjustment mechanism. It dynamically computes and adjusts the threshold based on error change trends during training, optimizing the client selection strategy. Secondly, we propose a dual-factor selection strategy, which adapts to different training stages by using distinct selection criteria. Finally, we propose a historical information interaction mechanism. It allows client selection to consider both current performance metrics and historical performance trends, enabling flexible adjustments. This mechanism significantly enhances training stability and convergence efficiency. Experimental results show that, compared to the Pow-d method, AHIBCS improves the global model accuracy by 2.1%, 2.4%, and 2.7% on the COVID-19, FMNIST, and CIFAR-10 datasets, respectively.