Reinforcement Learning-based Resource-Aware Adaptive Inference for Heterogeneous Split Federated Learning

Heterogeneous Split Federated Learning (HSFL) is a privacy-preserving technique for devices in different edge settings. However, the present methods adopt a static model partition that cannot adapt to network conditions, resulting in longer inference latency and higher communication costs. To address this issue, we propose an adaptive split inference approach for HSFL that can boost both model partitioning and communication efficiency in real time. We use a dynamic, resource-aware optimization mechanism to find the optimal split point, accounting for bandwidth and client computing power, coupled with a lightweight Reinforcement Learning (RL) policy that adapts to changing conditions. The validation in a multi-client simulation with varied bandwidths indicates that our solution reduces latency and overhead compared to static alternatives while maintaining model accuracy under varying network conditions, thereby enabling industrial Digital Twin (DT) synchronization and intelligent mmonitoring.