Adaptive UAV-Assisted Hierarchical Federated Learning: Optimizing Energy, Latency, and Resilience for Dynamic Smart IoT

Abstract: A key application of HFL lies in smart Internet of Things (IoT) systems, including remote monitoring and battlefield operations, where cellular connectivity is often unavailable. In such scenarios, UAVs can act as mobile aggregators, dynamically providing connectivity to terrestrial IoT devices. Subsequently, this paper investigates an HFL architecture enabled by energy-constrained, dynamically deployed UAVs that are susceptible to communication disruptions. We propose a novel approach to minimize global training costs in such environments by formulating a joint optimization problem that integrates learning configuration, bandwidth allocation, and IoT device-to-UAV association, ensuring timely global aggregation before UAV disconnections and redeployments. The problem explicitly captures the dynamic nature of IoT devices and their intermittent connectivity to UAVs and is shown to be NP-hard. To address its complexity, we decompose the problem into three interrelated subproblems. First, we optimize learning configuration and bandwidth allocation using an augmented Lagrangian function to reduce training costs. Second, we introduce a device fitness score that accounts for data heterogeneity (via Kullback-Leibler divergence), device-to-UAV proximity, and computational resources, leveraging a Twin Delayed Deep Deterministic Policy Gradient (TD3)-based algorithm for adaptive device-to-UAV assignment. Third, we develop a low-complexity two-stage greedy strategy for UAV redeployment and global aggregator selection, ensuring efficient model aggregation despite UAV disconnections.