Scholarlink

<p>High communication costs are a major challenge for Federated Learning (FL). Existing solutions often struggle to balance client availability and model accuracy while reducing communication overhead. In this paper, we propose Weighted Average Federated Learning via Two-Stage Client Sampling (WAFL-TSCS). Firstly, we use a detection signal to filter out available clients, enabling them to receive the global model and train the local model, thereby ensuring the efficiency and smoothness of the model training process. Then, we utilize the Kullback–Leibler divergence principle to evaluate the difference in weight distribution between the client’s local model and the global model, measuring the contribution and selecting clients with higher contributions to upload their model parameters. This approach reduces communication overhead. Finally, during the model parameter aggregation phase, the uploaded model parameters are weighted according to their contribution, achieving a weighted average update of the global model and improving model accuracy. Additionally, we introduce a dynamic adjustment term in the loss function to ensure algorithm convergence. Experimental results on two datasets, MNIST and CIFAR-10, show that WAFL-TSCS reduces communication overhead by 20% and improves model accuracy by over 0.05% compared to algorithms such as FedAvg and FedDM, demonstrating its effectiveness in reducing communication costs.</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

Wafl-tscs: an effective strategy to reduce communication costs in federated learning

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Debao Wang,
  • Shaopeng Guan,
  • Ruikang Sun

摘要

High communication costs are a major challenge for Federated Learning (FL). Existing solutions often struggle to balance client availability and model accuracy while reducing communication overhead. In this paper, we propose Weighted Average Federated Learning via Two-Stage Client Sampling (WAFL-TSCS). Firstly, we use a detection signal to filter out available clients, enabling them to receive the global model and train the local model, thereby ensuring the efficiency and smoothness of the model training process. Then, we utilize the Kullback–Leibler divergence principle to evaluate the difference in weight distribution between the client’s local model and the global model, measuring the contribution and selecting clients with higher contributions to upload their model parameters. This approach reduces communication overhead. Finally, during the model parameter aggregation phase, the uploaded model parameters are weighted according to their contribution, achieving a weighted average update of the global model and improving model accuracy. Additionally, we introduce a dynamic adjustment term in the loss function to ensure algorithm convergence. Experimental results on two datasets, MNIST and CIFAR-10, show that WAFL-TSCS reduces communication overhead by 20% and improves model accuracy by over 0.05% compared to algorithms such as FedAvg and FedDM, demonstrating its effectiveness in reducing communication costs.