Context-aware Online Federated Split Learning in 5G Networks
Shuang Wang, Wei Bao, Zhongming Huang, Binbin Ye, Jinshan Tang, Bo Xu · 2024
In order to meet the diverse needs of 5G intelligent applications, Federated Learning (FL) based on collaborative training across multiple devices can ensure data privacy while increasing the computational load for local model updates. On the other hand, Split Learning (SL), when properly implemented with model splitting, can adapt to the computing and transmission capabilities of different devices. This paper focuses on a Semi-decentralized Hybrid Federated Split Learning (SD-HFSL) framework that combines the advantages of FL and SL by breaking away from the limitations of a single central server and allowing shared split models to be aggregated among multiple edge servers. To maximize long-term training efficiency, we propose an online optimization problem that includes local model splitting and device association based on evaluated convergence performance. We also introduce a context-aware online training algorithm based on Contextual Multi-armed Bandit (CMAB) framework, where edge servers can observe context information at device sites for latency estimation and perform iterative optimization based on evaluated information in different contexts. Our experiments with multiple neural network models demonstrate significantly superior performance compared to other benchmarks.