Cluster-HSFL: A Cluster-Based Hybrid Split and Federated Learning

Federated learning (FL) and split learning (SL) are two machine learning paradigms that can distributedly train artificial intelligence (AI) models while guaranteeing client privacy. FL can parallelly train AI models based on clients’ dataset, but it suffers from considerable communication overhead for model uploading and imposes a significant on-device computation workload. SL can mitigate the computation workload of the devices by offloading part of the AI model training to an edge server, but it suffers from long training delay for a large number of clients. To address these challenges, in this paper we present a cluster-based hybrid split federated learning (Cluster-HSFL) scheme by integrating FL and SL paradigms. This scheme can reduce training latency and alleviate on-device computational constraints while preserving accuracy performance. We establish a testbed for the Cluster-HSFL scheme based on the NVIDIA Jetson series hardware. The experimental results demonstrate that the Cluster-HSFL surpasses FL and SL by 6.3× and 1.8×, respectively, in terms of delay performance. Our results provide insight into the practical performance of the Cluster-HSFL scheme and can help guide future experimental research on SL in resource-limited wireless networks.