Abstract: Federated learning(FL), as a privacy-preserving distributed learning framework, has effectively addressed the data isolation problem. However, the heterogeneous characteristics of IoT terminal devices in real-world scenarios, particularly performance variations and non-independent and identically distributed (Non-IID) data properties, were found to degrade model performance and convergence speed. To resolve these challenges, a federated split learning optimization method under edge heterogeneity named FedCRS(Federated Cluster-based Round Splitting) was proposed. First, an adaptive clustering strategy based on device capabilities was developed, through which clients were dynamically clustered and assigned customized sub-models to balance computational loads and mitigate straggler issues caused by device heterogeneity. Then, an innovative ring-topology cyclic model transfer mechanism was created, where local model fusion among intra-cluster clients was implemented to alleviate client drift from data heterogeneity while significantly enhancing global model robustness and generalization capability. Experimental evaluations on three label-heterogeneous datasets (FMNIST, CIFAR-10, and CIFAR-100) demonstrate superior performance of the proposed FedCRS method. Compared with five baseline methods (FedAvg, FedProx, MOON, SplitFed, and SplitMix), accuracy improvements of 8.7, 11.1, and 2.1 percentage

Key words: Keywords: federated learning, heterogeneity, split learning, cyclic topology, efficiency

摘要： 联邦学习（FL）作为一种以隐私保护为核心的分布式学习框架，已有效缓解了数据孤岛问题。然而，现实中的物 联网终端设备往往呈现出异构特性，特别是设备性能的差异以及数据的非独立同分布（Non-IID）特性，导致模型的性能和收 敛速度下降。针对这一问题，提出了一种边缘异构下的联邦分割学习优化方法 FedCRS（Federated Cluster-based Round Splitting）。 首先，FedCRS 通过基于设备性能的自适应聚类策略，动态地将客户端分簇，并为每个簇设计适配的子模型，以平衡计算负载， 解决设备异构性带来的落后者问题。其次，FedCRS 创新性地提出了一种环形拓扑的循环模型传递机制，通过簇内客户端局部 模型融合，缓解数据异构性带来的客户端漂移现象，并显著提升全局模型的鲁棒性与泛化能力。在 FMNIST、CIFAR-10 和 CIFAR-100 三个标签异构分布的数据集上的实验结果表明，所提出的 FedCRS 方法在精度上均取得最优表现，相比于 FedAvg、 FedProx、MOON、SplitFed 和 SplitMix 五种方法，至少提高了 8.7、11.1、2.1 个百分点，同时收敛速度至少提升了 73.7%、13.2% 和 17.3%。以上结果验证了 FedCRS 方法在面对边缘异构环境下优化联邦模型精度与收敛速度的有效性，展现出较好的实际应 用前景。

关键词: 联邦学习, 异构性, 分割学习, 循环拓扑, 高效性