A Lyapunov Optimization Enabled Energy-Efficient Dynamic Cut Layer Control Mechanism

This letter introduces a novel dynamic cut-layer control mechanism aimed at enabling energy-efficient training in the Split Federated Learning (SFL) process while satisfying specified latency and privacy constraints. The problem is formulated as a minimization of time-averaged energy consumption. The selection of the cut layer, which determines the split point dividing the model into client-side and server-side sub-models, plays a crucial role in balancing training workloads and enhancing the robustness of client-side models against data reconstruction attacks. To address this issue, our approach manages the cut layer dynamically, based on estimated total energy consumption and privacy requirements throughout the SFL process. We propose a theoretical framework to optimize the cut layer while maintaining the required round latency and privacy levels. This framework incorporates a virtual latency deficit queue and leverages Lyapunov optimization theory, transforming the long-term optimization problem into a series of round-wise drift-plus-cost minimization tasks. By dynamically minimizing the proposed cost function while ensuring latency and privacy constraints, the mechanism achieves effective optimization of SFL performance. Extensive simulations demonstrate that the proposed framework surpasses existing baselines by effectively reducing time-averaged energy consumption while meeting privacy and latency requirements. This is achieved through optimized cut layer control, ensuring stable training performance throughout.