Answers are grounded in this paper and cite their sources.
Autonomous vehicles rely on GNSS for lane-level localization, and GNSS spoofing is both cheap and well documented. Many detectors have been proposed, but they are seldom compared directly against an attacker that adapts to the detector, and evaluations often assume near-perfect localization. This paper reports a CARLA study of five detector families: a threshold rule, CUSUM, a Kalman chi-square innovation test, LSTM autoencoders, and statistical/learned hybrids. We test them against naive drift attacks and against a white-box, rate-limited stealthy attacker that greedily maximizes lateral displacement while staying under each detector's alarm. The localization backup is a realistic dead-reckoning odometry model with scale error, velocity noise, and heading-bias drift rather than ground truth. Over 2400 trials on 30 route-disjoint seeds, the simple Kalman chi-square test holds the adaptive attacker to 0.52 m of undetected displacement with no false alarms, compared with about 1.5 m for the rule and CUSUM. The LSTM detectors either raise too many false alarms to be usable or detect later than the chi-square test. Every detector eliminates collisions from a 4 m step attack, dropping the collision rate from 10% to 0%, and the ranking is stable across 0.5x to 2x odometry noise. It also holds on a second map and survives ablations of the braking gate, fail-safe, and odometry model. On 142 minutes of real highway driving (comma2k19), the same frozen detectors keep the ranking, with the chi-square test holding the stealthy attacker to 0.32 m. The scope is high-fidelity simulation with software-level sensor injection, cross-checked against real-world driving data.
No takes yet. Share an insight, caveat, or question.
Nikhil Yadav (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: