Joint Communication and Computing Resource Allocation for Energy Efficient Hierarchical Federated Learning in Marine Internet of Things

Federated learning has emerged as a promising approach for applications in the Maritime Internet of Things (M-IoT). However, its deployment in this context presents significant challenges, such as high communication costs due to frequent model exchanges between uncrewed surface vehicles (USVs) and ground based stations, as well as convergence issues arising from the non-independent and identically distributed (non-IID) nature of IoT data. To address these challenges, we propose a hierarchical federated learning model specifically designed for M-IoT systems, where high-altitude platforms (HAPs) are used as intermediaries to perform model aggregation in proximity to USVs. We jointly optimize the local training frequency, model aggregation frequency, as well as communication and computing resource allocation for HAPs and USVs, with the objective of minimizing total energy consumption while satisfying delay constraints. Due to the inherent non-convexity of the optimization problem, we decompose it into a top-level problem and a bottom-level problem, and develop a low-complexity alternating optimization algorithm to effectively solve the problem. Simulation results demonstrate that the proposed approach significantly reduces energy consumption compared to other benchmarks while satisfying the delay constraint.