Presentation Date: Feb 14, 2026
AGSA Abstract
Autonomous vehicles depend significantly on Global Navigation Satellite Systems (GNSS) for localization; nevertheless, these signals are susceptible to spoofing attacks that might jeopardize safety and dependability. Current research frequently tackles spoofing with either constrained real-world datasets or singular detection techniques, resulting in a deficiency of cohesive, reproducible frameworks that integrate several methodologies. This study presents a comprehensive framework for detecting GPS and Inertial Measurement Unit (IMU) spoofing, integrating both supervised and unsupervised learning methodologies. Synthetic GNSS–IMU trajectories were created with sudden shifts and slow drifts to emulate spoofing scenarios. Physics-informed attributes, such as GNSS–INS residuals, displacement statistics, and satellite signal fluctuations, were retrieved to facilitate model training. Supervised ensemble classifiers (Random Forest and XGBoost) were assessed using an 80:20 split and stratified K-fold cross-validation, attaining F1-scores exceeding 0.95 and ROC-AUC values approaching 1.0. An unsupervised LSTM Autoencoder was used to detect anomalies using reconstruction error in order to address the lack of labeled data. Results indicate that ensemble classifiers deliver highly dependable supervised detection, whereas the autoencoder provides supplementary functionality in scenarios with limited labeled data. The suggested pipeline establishes a reproducible methodology that can be extended to large-scale datasets, thereby enhancing spoofing resilience in autonomous car systems.
No Co Authors Found
