Federated Learning via Non-Orthogonal Multiple Access for UAV-Assisted Internet of Things

Federated learning (FL), utilizing data from edge devices (EDs) while protecting user privacy, has gained much attention. Its efficacy is substantially influenced both by the quantity of connected devices and the quality of wireless communications. Network congestion, resulting from multiple access and signal attenuation caused by physical obstacles may severely impact the convergence of the FL model. To address these issues, this paper employs non-orthogonal multiple access (NOMA) for uplink transmission and designs a two-tier FL framework consisting of ground devices and unmanned aerial vehicles (UAVs) to ensure the construction of line-of-sight (LoS) channels from EDs to the base station. Moreover, we construct a multi-objective joint optimization problem to minimize the FL convergence time considering constraints such as the NOMA uplink latency, ED selection strategy, local training latency, and energy consumption. We also deduce the theoretical upper bound of the convergence time and transform the proposed multi-objective problem into a solvable form by eliminating the discrete variables determined by ED selection. In turn, we utilize the proximal policy optimization (PPO) algorithm to solve this optimization problem. Finally, extensive experimental results demonstrate the advantages of our proposed algorithm in terms of latency and energy consumption, while yielding a high robustness and scalability.