Computational offloading and delay minimization for UAV-aided edge federated learning

Abstract

Future wireless networks are expected to support increasingly diverse, data-intensive, and latency-critical applications driven by massive deployments of Internet of things (IoT) devices. As dynamic traffic grows across next-generation wireless systems, efficient resource optimization becomes essential to ensure reliability, scalability, and timely decision-making. Traditional optimization techniques, while mathematically rigorous, often suffer from high computational complexity and limited adaptability in large-scale heterogeneous environments. In contrast, machine learning has emerged as a powerful alternative for resource optimization in future wireless networks. This thesis leverages insights from recent survey findings on ML-based resource optimization, particularly within federated learning (FL) environments, to design adaptable optimization frameworks suited for dynamic wireless systems. Building upon this foundation, the thesis presents three increasingly capable uncrewed aerial vehicle (UAV)-assisted FL frameworks. First, we develop a UAV-aided FL (UAFL) system that mitigates the straggler effect by offloading partial datasets from resourceconstrained IoT devices to UAV-mounted mobile edge computing (MEC) servers. We formulate a system-delay minimization problem under computation, communication, and quality of service (QoS) constraints, solving it using epigraph transformation, deterministic simplex optimization, and deep reinforcement learning (DRL) for improved run-time performance. Next, we propose a hierarchical queue-based UAV-aided FL framework. The system introduces multi-queue architectures at IoT devices, follower UAVs, and leader UAVs to manage irregular data arrivals and heterogeneous processing capabilities. We apply Lyapunov drift-plus-penalty optimization to stabilize queues and minimize delay by decomposing the problem into sequential quadratic programming (SQP)-solvable subproblems. Simulation results demonstrate significant improvements in delay reduction, queue stability, and overall system throughput compared to standard FL and UAFL. Finally, we extend the hierarchical system by integrating generative adversarial networks (GANs) to counter the impact of non-independent and identical data (Non-IID) data across distributed devices. Offloaded datasets are utilized to train distributed GANs at follower UAVs, while aggregated generator models at the leader UAV produce synthetic samples that enhance data diversity. This GAN-augmented HUAFL framework improves model accuracy, accelerates convergence, and maintains queue stability within a UAV-MEC-assisted architecture. Experimental results show that this integrated approach outperforms baseline FL, hierarchical FL, and UAFL across accuracy, latency, and resource utilization metrics. Overall, this thesis presents a unified, scalable, and computation-efficient set of frameworks for FL in mission-critical IoT environments. By integrating ML-based resource optimization, hierarchical queuing, and GAN-based data augmentation, the proposed approaches advance the state of UAV-assisted edge intelligence for next-generation wireless networks.