HGRL-ATN: a hierarchical graph reinforcement learning framework for trustworthy UAV navigation under GPS spoofing attacks

Abstract Unmanned aerial vehicles (UAVs) that rely on the Global Positioning System (GPS) for localization are increasingly exposed to spoofing attacks that induce large positional deviations and mission failure. Established defense model-based estimators, Extended Kalman Filter (EKF) sensor fusion, and standard deep reinforcement learning (DRL)-are largely reactive, generalize poorly to unseen attack patterns, and do not model the relational structure that links heterogeneous sensor modalities. This paper presents HGRL-ATN, a framework that couples graph-neural-network (GNN) sensor fusion, variational trust estimation, and hierarchical reinforcement learning (HRL) into a single end-to-end pipeline for secure UAV navigation. Sensors are represented as nodes of a graph and fused by a graph attention network, allowing cross-modal inconsistencies induced by spoofing to be detected relationally. A variational module assigns per-sensor trust and uncertainty scores that down-weight compromised inputs, while a two-level policy separates navigation strategy selection from continuous control. A digital-twin module predicts next-state dynamics to support anticipatory decision making. Evaluated on a synthetic UAV dataset, the VeReMi benchmark, and CARLA-generated scenarios, HGRL-ATN attains a detection accuracy of 97.0\%, an F1-score of 0.965, and a navigation success rate of 94.8\%, against 82.3\%--90.4\% for representative baselines. Ablation confirms that graph fusion and trust estimation are the dominant contributors, and trajectory analysis shows substantially reduced path deviation under active spoofing.