JaGuard: Jamming Correction of GNSS Deviation with Deep Temporal Graphs

Abstract

Global Navigation Satellite Systems (GNSS) are increasingly exposed to intentional jamming, threatening reliability when accurate positioning and timing are most critical. We address this problem by formulating interference mitigation as a dynamic graph regression task and propose JaGuard, a receiver-centric temporal graph neural network that estimates and corrects latitude and longitude errors. At each 1 Hz epoch, the satellite-receiver scene is represented as a heterogeneous star graph with time-varying satellite attributes such as SNR, azimuth and elevation. A single-layer HeteroGCLSTM fuses one-hop spatial context with short-term temporal dynamics to produce a 2D deviation estimate. We evaluate JaGuard on data collected from two commercial receivers under controlled conducted jamming using three jammer types (CW, 3xCW, FM) and six power levels from -45 to -70 dBm, each repeated 50 times across pre-jam, jam, and recovery phases. JaGuard outperforms strong multivariate baselines (TSMixer, uniform CNN, Seq2Point) in all conditions. Under severe jamming at -45 dBm, it achieves 3.64-7.74 cm MAE, improving to 1.59-1.90 cm for -60 to -70 dBm. On mixed-mode datasets, it attains 3.78 cm MAE on GP01 and 4.25 cm on U-blox 10. With only 10 percent of the training data, JaGuard remains ahead, reaching about 20 cm MAE compared to 36-42 cm for the baselines.