GNSS-Spoofing Detection via Machine Learning with In-Domain and Cross-Domain Testing and Open-Access I/Q Data

Abstract

Global Navigation Satellite Systems (GNSS) are essential for applications that rely on positioning, velocity, and timing (PVT) information. Unfortunately, a surge in interferences in GNSS frequency bands causes severe security risks for various navigation-based applications. Traditional signal-processing techniques cannot generalize well to different interference types, therefore, Radio Frequency Fingerprinting (RFF) methods using sophisticated Machine Learning (ML) methods are used to capture complex patterns in GNSS signals. However, ML-based spoofing detection studies have relied primarily on the testing of the ML models with the same type of spoofing signals as those used in the training phase (the in-domain testing). This is a strong limitation of ML-based RFF, as real-time spoofers are highly variable and time-varying. Therefore, the ability of an ML algorithm to provide cross-domain generalization across unseen datasets is crucial. To this end, our paper presents a comprehensive analysis of achievable spoofing-detection accuracies using four ML models and compares the in-domain with cross-domain testing accuracies. We used multiple scenarios from two large open-access GNSS spoofing repositories: the Texas Spoofing Test Battery (TEXBAT) and the Oakridge Spoofing Test Battery (OAKBAT). For in-domain testing with mixed sample split, all ML models achieve an average classification accuracy greater than 90%. The accuracy for PRN-level becomes slightly lower, with around 80% in GPS L1 and around 75% in Galileo E1. Finally, the average accuracy in the cross-testing approach drops to 66% or less, highlighting the need for more advanced ML models to achieve higher generalization capabilities under complex spoofing conditions.