Accelerating Wireless Hierarchical Split Federated Learning via Smashed Data Generation

Abstract

As the size of machine learning models scales up, conventional federated learning (FL) encounters challenges on wireless edge clients due to scarce resources. Split federated learning (SFL) mitigates these constraints by offloading part of the model to the edge server. However, most prior work adopts two-tier synchronous SFL, which suffers from stragglers arising from heterogeneous computing and communication, and a single edge server becomes a bottleneck as participation scales. We propose SAGE-HSFL, a novel smashed data generating-enabled semi-asynchronous hierarchical split federated learning framework. Specifically, SAGE generates synthetic smashed data from the empirical distribution of historical activations to reduce bias from data heterogeneity. We derive a convergence bound for the framework and show that SAGE reduces bounded dissimilarity induced by stragglers. Inspired by this, we formulate a joint client scheduling, cut layer decision, and bandwidth allocation optimization problem, and decompose it into tractable subproblems using Lyapunov optimization. We then design algorithms for bandwidth allocation, cut layer decision, and dynamic client scheduling. Experiments on FMNIST and CIFAR-10 show faster convergence and higher test accuracy than representative baselines.