Split Federated Learning (SFL) trains large deep models on resource‑constrained edge devices by partitioning the network between client and server. Asynchronous SFL lets each client proceed independently, avoiding the straggler problem caused by system heterogeneity, but it introduces two coupled inconsistencies. Horizontal inconsistency refers to divergent activation sequences across clients, while vertical inconsistency denotes mismatched parameters between clients and the server. Statistical heterogeneity amplifies these effects, leading to biased activations, model divergence, and severe client drift.
CoASFL addresses these issues with an inconsistency‑aware asynchronous SFL framework. It features a dynamic activation merging mechanism that aligns and fuses activation streams from heterogeneous clients on the server, producing more consistent activations and mitigating horizontal inconsistency. A consistency‑aware model aggregation mechanism adjusts aggregation weights based on client participation frequency and progress gaps, reducing vertical inconsistency. Joint coordination of activation‑level and model‑level consistency suppresses inconsistency propagation during asynchronous training.
Experiments on an 80‑device heterogeneous edge testbed show that CoASFL consistently outperforms state‑of‑the‑art SFL approaches, achieving 1.61×~3.74× faster training and up to 3.26% higher final model accuracy.
Review: By jointly handling horizontal and vertical inconsistencies, CoASFL markedly improves the efficiency and robustness of asynchronous split federated learning in heterogeneous settings.