Wireless Distributed Learning: A New Hybrid Split and Federated Learning Approach

Cellular-connected unmanned aerial vehicle (UAV) with flexible deployment is foreseen to be a major part of the sixth generation (6G) networks. The UAVs connected base station (BS), as users (UEs), could exploit machine learning (ML) algorithms provide wide range advanced applications, like object detection and video tracking. Conventionally, ML model training performed at BS, known centralized (CL), which causes high communication overhead due transmission large datasets, potential concerns about UE privacy. To address this, distributed algorithms, including federated (FL) split (SL), were proposed train models in manner via only sharing parameters. FL requires higher computational resource on side than SL, while SL has larger when local dataset large. effectively an considering diversity UEs different capabilities channel conditions, we first propose novel architecture, hybrid (HSFL) algorithm by reaping parallel mechanism splitting structure SL. We then its convergence analysis under non-independent identically (non-IID) data random selection scheme. By conducting experiments two models, Net AlexNet, wireless UAV networks, our results demonstrate that HSFL achieves accuracy less IID non-IID data, increases increasing number UEs. further Multi-Arm Bandit (MAB) based best (BC) 2-norm (BN2) (MAB-BC-BN2) scheme select better quality updates for each round. Numerical it BC, MAB-BC MAB-BN2 non-IID, Dirichlet-nonIID Dirichlet-Imbalanced data.