C²-SFL:Class-Balanced and Cost-Aware Split Federated Learning for Mobile Edge Computing

Split Federated Learning (SFL) facilitates collaborative training in Mobile Edge Computing (MEC) by splitting models between clients and servers. However, existing client selection and resource allocation strategies fail to account for data and system heterogeneity, thereby exacerbating data skew and leading to inefficient training and high costs. To address these issues, we propose a Class-Balanced and Cost-Aware SFL (C²-SFL) framework that jointly optimizes client selection, model partitioning, and bandwidth allocation. We introduce a Relation-Aware Class-Balanced Algorithm (RACBA) to select clients by balancing data distribution and energy states, along with a Cost-Driven Optimization Algorithm (CDOA) to dynamically adjust model split points and bandwidth. Experiments show that C²-SFL significantly reduces training costs, accelerates convergence, and improves accuracy under heterogeneous MEC settings, achieving reductions of 12.6%-71.5% in energy consumption and 5.2%-50.9% in training latency.