Enhancing Cross-Scenario GNSS Spoofing Detection with WGAN-GP Data Augmentation

Machine-learning-based GNSS spoofing detectors are often trained on limited and highly scenario-dependent datasets, which severely degrades their generalization performance and leads to high false alarm rates in unseen environments. In this paper, we revisit GNSS spoofing detection from a data-centric perspective and propose a WGAN-GP–based augmentation framework tailored to satellite navigation signals. Using the TEXBAT spoofing dataset, we extract 82-dimensional SQM features from a software receiver and study the challenging cross-scenario setting where a CNN detector is trained on one spoofing scenario and tested on a different one. To improve robustness under small-sample conditions, we construct synthetic navigation signals with both classical GAN and WGAN-GP models, and plug the generated data into a standard CNN detector without modifying its architecture. We further introduce an evaluation protocol that fixes the false positive rate (FPR) at practical operating points (0.01, 0.05, and 0.1) and compares detection rate and accuracy across the original and augmented datasets. Experiments show that GAN-based augmentation already provides significant gains over the original TEXBAT scenario, while WGAN-GP consistently outperforms GAN, achieving about a 5% higher detection rate and 2.5% higher accuracy at the same FPR. Moreover, WGAN-GP produces roughly one order of magnitude more usable samples per generation with much more stable training dynamics. These findings indicate that WGAN-GP is a more suitable augmentation tool for small-sample GNSS spoofing data and provide concrete guidelines for deploying data-driven spoofing detectors in heterogeneous real-world environments.