Spatio-Temporal Graph Neural Network for GNSS Spoofing Detection: A Cross-Scenario Generalization Study

Abstract

Global Navigation Satellite System (GNSS) spoofing attacks use counterfeit signals to manipulate a receiver's position, velocity, and time, threatening safety-critical applications. Conventional detection techniques use indicators such as the carrier-to-noise density ratio and correlation-function distortion, but advanced spoofers can match authentic signal power and preserve nominal correlation shapes, reducing their effectiveness. Deep learning detectors can identify patterns missed by these conventional techniques, but they often model satellite channels independently, and their generalization to unseen attacks remains insufficiently evaluated. We therefore propose a spatio-temporal graph neural network (ST-GNN) that represents the satellite constellation as a dynamic graph. Nodes contain tracking-loop observables, while edges encode normalized pairwise differences in these observables. We evaluate the model on the Texas Spoofing Test Battery (TEXBAT) using a mixed-split protocol and two cross-scenario protocols that progressively exclude attack categories from training. Under mixed-split evaluation, the ST-GNN achieves a detection rate of 96.4% and a false alarm rate of 2.26%, attaining the highest detection rate among the compared methods. When ds6 and ds8 are excluded from training while ds7 is retained, the overall detection rate decreases to 58.6%. Detection remains high for the held-out security code estimation and replay (SCER) variant but deteriorates substantially for dynamic matched-power spoofing, and removing all SCER variants from training causes further degradation. The findings demonstrate the value of inter-satellite modeling for known and related attacks, but generalizing to entirely unseen spoofing categories remains an challenge.