Detecting GPS Spoofing Attacks with a Trust Index Using Sensor Cross-Validation Based on Inertial Information

GPS spoofing poses a critical threat to autonomous mobile robots that rely on unencrypted civilian GPS signals for navigation. Although the PX4 autopilot’s extended Kalman filter (EKF2) incorporates innovation gating to reject anomalous measurements, spoofed signals that emulate legitimate GPS characteristics with low reported error metrics can bypass this defense. This paper presents a sensor cross-validation framework that detects GPS spoofing by cross-validating the reported GPS solution against the inertially propagated state estimate. Eight heuristic detection rules are evaluated as a training-free baseline, while the Random Forest classifier primarily exploits GPS accuracy and EKF2 innovation features. Ground-truth annotations for training and evaluation are taken from the simulator’s known spoofing timelines across 60 synthetic flights, independently of the heuristic rules themselves. A Random Forest classifier trained on 30-sample sliding windows (15-sample stride) across 34 telemetry fields achieves 99.8% classification accuracy with a spoofed-class F1-score of 0.997 on a dataset of 7920 labeled windows collected from 60 simulated flights across three terrain types in the PX4 SITL–Gazebo environment under varied wind and spoofing conditions. The proposed GPS Trust Index provides a continuous real-time integrity metric complementary to the binary classifier output. The simulation results demonstrate the potential of lightweight machine learning on existing autopilot telemetry to augment EKF2-based spoofing detection, without requiring additional sensor hardware.