Multi-Tier Split Federated Learning for Multi-Level Personalization

Achieving multi-level personalization in federated learning (FL) is increasingly important in multi-tier computing networks, where client preferences are influenced not only by individual data but also by regional environments. However, existing personalized FL approaches focus primarily on userspecific data and fail to capture this hierarchical structure. We propose Multi-Tier Split Federated Learning (MT-SFL), an efficient FL framework that enables multi-level personalization by integrating split learning with a multi-exit model architecture. MT-SFL performs both training and inference across different network tiers, reflecting heterogeneous computing capabilities and significantly reducing client-side computation, communication overhead, and inference latency. Simulation results demonstrate that MT-SFL effectively personalizes models to clientpreferred tasks while capturing broader regional and global preferences, providing a scalable and resource-adaptive solution for hierarchical FL systems.