Split federated learning-driven resource-efficient MEC framework for UAV-based networks

Distributed collaborative machine learning techniques enable the training of intelligent models while preserving user data privacy. However, in reality, training a large-scale and intricate model on resource-constrained devices such as Unmanned Aerial Vehicles (UAVs) is unfeasible. In this context, lightweight and resource-efficient deep learning techniques are required. This work first suggests a new resource-aware distributed framework, SFMec, in the context of a UAV power consumption scenario. The framework is evaluated and compared with other distributed frameworks, including FedMec, a federated learning-based approach, to assess its performance across different system architectures and resource management strategies. The results obtained demonstrate that SFMec has the potential to conserve more than 50% of the storage space occupied by FedMec, making it more attractive for devices with limited resources. Then, a novel architecture, denoted as SFMecLite, is introduced to minimize the interactions between SFMec entities. Furthermore, an enhanced version of SFMecLite is also presented that greatly outperforms FedMec and reduces the computational and communication costs in SFMec without compromising learning performance.