Scholarlink

Federated learning (FL) facilitates collaborative model training while preserving data privacy. However, it faces challenges due to data heterogeneity, which adversely affects performance and efficiency. Clustered Federated Learning (CFL) addresses this issue by grouping clients with similar data distributions. Nevertheless, traditional CFL weakens parameter correlation during aggregation (parameter incoherence), thereby limiting further enhancements. Motivated by the benefits of serial training in maintaining parameter coherence, this paper introduces a novel FL framework, FedCD, which integrates serial training into a parallel FL environment to mitigate the impact of data heterogeneity. FedCD incorporates two key strategies: Cascaded Optimization (CO) and Degraded Clustering (DC). The CO strategy embeds the serial training into each client cluster to enhance parameter coherence while preserving inter-cluster parallelism for efficiency. However, the CO-strategic integration of serial and parallel training creates an optimization conflict, whose performance depends heavily on the consensus of intra-cluster models. Therefore, we further introduce the DC strategy, which employs a degradation mechanism for client clustering—adaptively adjusting the cluster numbers and gradually reducing the priority of parallel training during FL. The DC strategy prioritizes parallel training to ensure efficiency in the early stages of FL when intra-cluster consensus is limited. The preference adaptively transitions to serial training within fewer clusters as updates stabilize, thus enhancing both efficiency and overall performance. Our experiments on both synthetic and real-world data demonstrate FedCD’s effectiveness in balancing efficiency and accuracy in FL.

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

FedCD: A Hybrid Federated Learning Framework for Adaptive Training Under Data Heterogeneity

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Weide Zhan,
  • Baoyao Yang,
  • Zhixiang Fang,
  • Dongzhe Li,
  • Yali Ma,
  • Yiqun Zhang

摘要

Federated learning (FL) facilitates collaborative model training while preserving data privacy. However, it faces challenges due to data heterogeneity, which adversely affects performance and efficiency. Clustered Federated Learning (CFL) addresses this issue by grouping clients with similar data distributions. Nevertheless, traditional CFL weakens parameter correlation during aggregation (parameter incoherence), thereby limiting further enhancements. Motivated by the benefits of serial training in maintaining parameter coherence, this paper introduces a novel FL framework, FedCD, which integrates serial training into a parallel FL environment to mitigate the impact of data heterogeneity. FedCD incorporates two key strategies: Cascaded Optimization (CO) and Degraded Clustering (DC). The CO strategy embeds the serial training into each client cluster to enhance parameter coherence while preserving inter-cluster parallelism for efficiency. However, the CO-strategic integration of serial and parallel training creates an optimization conflict, whose performance depends heavily on the consensus of intra-cluster models. Therefore, we further introduce the DC strategy, which employs a degradation mechanism for client clustering—adaptively adjusting the cluster numbers and gradually reducing the priority of parallel training during FL. The DC strategy prioritizes parallel training to ensure efficiency in the early stages of FL when intra-cluster consensus is limited. The preference adaptively transitions to serial training within fewer clusters as updates stabilize, thus enhancing both efficiency and overall performance. Our experiments on both synthetic and real-world data demonstrate FedCD’s effectiveness in balancing efficiency and accuracy in FL.