Stitch-able Split Learning Assisted Multi-UAV Systems

Unmanned aerial vehicles (UAVs), commonly known as drones, have gained widespread popularity due to their ease of deployment and high agility in various applications. In scenarios such as search missions and target tracking, conducting complex and computation-intensive tasks in multi-UAV systems have become essential. Recent investigations have explored the integration of collaborative centralized learning (CL) and federated learning (FL) into multi-UAV systems. However, CL methods raise privacy concerns and may suffer from communication delays, while FL methods demand high UAV-side computation capability. To address these challenges, split learning (SL) emerges as a promising alternative, offering reduced learning iteration time and improved accuracy in resource-constrained edge clients. In this study, we leverage SL and Stitch-able Neural Network (SN-NET) to propose a novel Stitch-able Split Learning (SSL) approach for multi-UAV systems. The proposed SSL approach is capable of tackling challenges in terms of device instability and model heterogeneity that associated in multi-UAV systems. Comparative simulations are conducted, evaluating its performance against CL, FL, traditional SL and SFLV1 (SplitFed Learning V1) approaches to establish its superiority.