Scholarlink

Federated learning (FL) has been considered a promising paradigm for decentralized model training. However, its practical implementation is hampered by system heterogeneity clients vary significantly in computing power, memory, and network conditions and statistical heterogeneity due to non-IID data. In this paper, we propose an FL framework that simultaneously addresses both challenges through a two-stage approach. First, clients are clustered based on hardware configuration into performance-based groups. Second, each cluster is assigned a separate FL algorithm that matches the statistical heterogeneity of that cluster. Furthermore, we adjust the number of local epochs and the aggregation frequency per cluster to harmonize the training progress across the system. We validate our framework using the CIFAR-10 dataset with ResNet-18 as the global model. Experimental results show that our method achieves 92.34% accuracy on the global model while maintaining effective participation across all clusters, demonstrating the effectiveness of the choice of training adaptation and conditioning algorithms.

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

Scalable and Fair Federated Learning for Heterogeneous Devices via Cluster-Based Optimization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hoang Quang Anh,
  • Dao Thi Thanh,
  • Bui Dinh Chien,
  • Tran Van Viet,
  • Nguyen Dinh Manh Linh

摘要

Federated learning (FL) has been considered a promising paradigm for decentralized model training. However, its practical implementation is hampered by system heterogeneity clients vary significantly in computing power, memory, and network conditions and statistical heterogeneity due to non-IID data. In this paper, we propose an FL framework that simultaneously addresses both challenges through a two-stage approach. First, clients are clustered based on hardware configuration into performance-based groups. Second, each cluster is assigned a separate FL algorithm that matches the statistical heterogeneity of that cluster. Furthermore, we adjust the number of local epochs and the aggregation frequency per cluster to harmonize the training progress across the system. We validate our framework using the CIFAR-10 dataset with ResNet-18 as the global model. Experimental results show that our method achieves 92.34% accuracy on the global model while maintaining effective participation across all clusters, demonstrating the effectiveness of the choice of training adaptation and conditioning algorithms.