Scholarlink

The rapid growth of Internet of Things (IoT) networks introduces challenges in processing distributed and heterogeneous data. Federated Learning (FL) addresses these by enabling collaborative model training without compromising data privacy, but the non-IID nature of the IoT data hampers the performance and convergence of the model. Clustered FL has the potential to address this dilemma by grouping similar clients into a cluster. However, existing clustered FL methods are either static or particularly complicated in the training process. To tackle this, we propose an efficient Dynamic Spectral Clustered FL (DSCFL) framework, which groups clients adaptively based on data similarity using spectral clustering. This approach reduces the impact of data heterogeneity, improving local and global model performance. Experiments on different datasets show that DSCFL outperforms existing methods in accuracy, scalability, and robustness, providing a robust framework for FL in commercial applications.

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

Enhancing Federated Learning in IoT Through Dynamic Spectral Clustering

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Changyu Zheng,
  • Junli Gong,
  • Jianxiong Guo,
  • Zhiqing Tang,
  • Tian Wang

摘要

The rapid growth of Internet of Things (IoT) networks introduces challenges in processing distributed and heterogeneous data. Federated Learning (FL) addresses these by enabling collaborative model training without compromising data privacy, but the non-IID nature of the IoT data hampers the performance and convergence of the model. Clustered FL has the potential to address this dilemma by grouping similar clients into a cluster. However, existing clustered FL methods are either static or particularly complicated in the training process. To tackle this, we propose an efficient Dynamic Spectral Clustered FL (DSCFL) framework, which groups clients adaptively based on data similarity using spectral clustering. This approach reduces the impact of data heterogeneity, improving local and global model performance. Experiments on different datasets show that DSCFL outperforms existing methods in accuracy, scalability, and robustness, providing a robust framework for FL in commercial applications.