Online Node Selection Strategy for SplitFed Learning in Mobile Communication
Boyu Jin, Hui Tian, Shaoshuai Fan · 2023
In the era of 5G and B5G, the deep integration of AI with mobile communications has become increasingly apparent, intensifying the urgency of deploying learning architectures in resource-constrained mobile environments. SplitFed Learning (SFL) has gained prominence due to its advantages in conserving communication overhead and computational resources while ensuring privacy preservation. It is crucial to address the challenges posed by dynamic network conditions and stringent resource constraints. This paper presents a deployment method for SFL in mobile edge situations, focusing on the investigation of the node selection problem within dynamic network structures. The proposed online learning algorithm based on Lyapunov stochastic optimization theory allows for the approximation of optimal node decisions without relying on prior knowledge, thereby constraining communication costs and enhancing training outcomes. Simulation results reveal that our approach significantly reduces communication overhead compared to benchmark methods while maintaining or even enhancing learning performance.