Impact of neural network depth on split federated learning performance in low-resource UAV networks

Training without sharing data is one of the drivers that makes Federated Learning (FL) more attractive, compared to centralized approaches. However, requiring each learner to train the full model may not be efficient, particularly for devices with restricted resources, such as those available in Unmanned Aerial Vehicles (UAVs). To address this issue, a variation of FL technique, specifically Split Federated Learning (SFL), has recently been proposed. Unlike FL, the key concept of SFL is to divide the layers of the neural network among the involved learners. Therefore, each individual client will train only a segment of the model (submodel) rather than the entire model. Clearly, this technique, besides data privacy, optimizes the utilization of computational resources, reduces client-side training time, and enhances model privacy. However, there are questions that require answers: How should we split the model? Shall we systematically divide it in half, or is there a more optimal approach? In this line of thought, this paper provides a detailed analysis of possible splitting schemes of a power consumption prediction model for UAVs. First, the SFL-enabled model is presented. Second, an experimental analysis is conducted in which different splitting alternatives are made and numerically analyzed to examine the influence of network layering on split federated learning performance.