Semi-Asynchronous Federated Split Learning for Computing-Limited Devices in Wireless Networks

The rapid evolution of edge computing and artificial intelligence (AI) paves the way for pervasive intelligence in the next-generation network. As a hybrid training paradigm, federated split learning (FSL) leverages data and model parallelism to enhance training efficiency. However, existing FSL encounters unacceptable waiting latency due to device heterogeneity and synchronous model aggregation. To address this issue, we propose a semi-asynchronous FSL (SAFSL) framework that enables personalized model splitting and aperiodic model aggregation. We derive the convergence upper bound by considering factors such as the number of devices, training iterations, and data heterogeneity. To minimize the long-term average training latency while maintaining high energy efficiency in resource-constrained wireless networks, we formulate a stochastic mixed-integer nonlinear programming problem. By decomposing it into multiple sub-problems in each round, we propose a Lyapunov-based alternating optimization algorithm to solve it in an online manner. Numerical results demonstrate that our SAFSL achieves faster convergence with reduced communication overhead while maintaining high prediction performance under non-independent and identically distributed data, outperforming state-of-the-art benchmarks. Moreover, our algorithm achieves a low training latency, highlighting its superior performance and effectiveness.