Scholarlink

Federated learning (FL) has emerged as a prominent paradi-gm for distributed artificial intelligence, enabling collaborative model training without compromising data privacy. However, conventional FL mhods often suffer from performance degradation under non-IID data distributions. While clustered FL (CFL) mitigates this issue to some extent, it still faces challenges in ensuring robust clustering quality and convergence efficiency, particularly with large-scale clients, high-dimensional features, or noisy data. To overcome these limitations, we propose EnvCl-uster-FL, an environment-aware federated learning framework. This approach begins by assessing four environmental factors: data distribution, client scale, feature dimensionality, and noise level. Based on this assessment, it dynamically selects a suitable clustering strategy. EnvCluster-FL replaces standard cosine similarity with a radial basis function (RBF) kernel-based similarity measure and adopts a two-tier aggregation architecture comprising a global representation layer and an intra-cluster mapping layer. Additionally, a clustering-segmentation stopping criterion is introduced to regulate the aggregation and segmentation processes. Experimental results indicate that EnvCluster-FL achieves more accurate client grouping and faster convergence compared with baseline methods.

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

EnvCluster-FL: An Environment-Aware Clustering Federated Learning Framework

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chuanyu Huang,
  • Fanli Sun,
  • Zhongxiang Shi,
  • Junwei Du,
  • Feng Jiang

摘要

Federated learning (FL) has emerged as a prominent paradi-gm for distributed artificial intelligence, enabling collaborative model training without compromising data privacy. However, conventional FL mhods often suffer from performance degradation under non-IID data distributions. While clustered FL (CFL) mitigates this issue to some extent, it still faces challenges in ensuring robust clustering quality and convergence efficiency, particularly with large-scale clients, high-dimensional features, or noisy data. To overcome these limitations, we propose EnvCl-uster-FL, an environment-aware federated learning framework. This approach begins by assessing four environmental factors: data distribution, client scale, feature dimensionality, and noise level. Based on this assessment, it dynamically selects a suitable clustering strategy. EnvCluster-FL replaces standard cosine similarity with a radial basis function (RBF) kernel-based similarity measure and adopts a two-tier aggregation architecture comprising a global representation layer and an intra-cluster mapping layer. Additionally, a clustering-segmentation stopping criterion is introduced to regulate the aggregation and segmentation processes. Experimental results indicate that EnvCluster-FL achieves more accurate client grouping and faster convergence compared with baseline methods.