On Cross-Testing Datasets for RF-Fingerprinting Based Deep-Learning GNSS Spoofing Detection

Abstract

Global Navigation Satellite Systems (GNSS) are crucial for acquiring position, velocity, and time (PVT) information for various applications such as air navigation, precision agriculture, logistics, and emergency services. GNSS is a backbone technology that supports many crucial aspects of modern-day life and industry. Unfortunately, GNSS is currently facing significant security risks due to a surge in signal interference from jammers and spoofers (i.e., fake GNSS signals), which can render it unreliable. To address this ever-growing challenge, novel GNSS receiver algorithms are currently being developed using machine learning (ML). These are fully reliant on available GNSS spoofing repositories such as TEXBAT and OAKBAT. However, in the current state-of-the-art, there are still inconsistencies on how models are trained using a single dataset and cross-tested with others. To address this issue, our work offers an overview of the impact of cross-testing with different datasets and demonstrates that results vary significantly depending on the dataset used for training and that with proper training datasets, cross-testing accuracy above 90 % can be achieved.