Optimizing Efficient Federated Split Learning Over Wireless Networks: Challenges and Opportunities

Federated learning (FL) has emerged as a promising solution over wireless networks, enabling a larger number of devices to jointly train an accurate deep learning (DL) model without collecting local data. However, considering the DL model with an increasing number of parameters and extremely high computing costs, a part of resource-constrained devices can hardly complete local model training. To solve this issue, since split learning (SL) can reduce the computing load of devices by splitting the DL model, we call for the development of SL to FL and focus on the federated split learning (FSL) framework. We discuss the fundamental research challenges of existing FSL and propose a novel FSL framework that utilizes a decentralized structure and adopts submodular optimization to improve the training efficiency with lower complexity in each training round. We further give a case study of the proposed framework in space-air-ground integrated networks (SAGINs), considering model splitting, device association, and resource allocation. Experiments are conducted to verify the effectiveness of the proposed framework with heterogeneous devices, where the convergence speed in the time dimension is increased by 15%. Finally, several future research directions identified in FSL are discussed.