Optimal Model Partition with Privacy Protection for Split Federated Learning in Wireless Mobile Networks

Split Federated Learning (SFL) is a distributed machine learning approach combining the strengths of federated learning and split learning. It not only effectively balances the computing resources between mobile devices and servers, reducing the training burden on resource-constrained mobile terminals, but also enhances data privacy protection during collaborative model training. However, determining the optimal partition point of the model is nontrivial, as it directly impacts communication overhead and privacy risks. In this paper, we propose an enhanced reinforcement learning-based partition-point selection scheme that effectively minimizes both communication costs and privacy risks for SFL in wireless mobile networks. Experimental results demonstrate the effectiveness and efficiency of the proposed approach.