AI Models Classify Driving Systems, Vulnerable to Telematics Jitter

Automated driving systems (ADSs) are becoming increasingly common, and future software-defined vehicles may run multiple ADSs. It is crucial to have independent monitoring systems to verify which ADS is active for safety, compliance, and anomaly detection. This research investigated the effectiveness of three sequence-based classification models—Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and a Transformer encoder model—in identifying Level 2 ADSs (Comma Openpilot, Tesla Autopilot, Cadillac Super Cruise) and manual driving using only vehicle telematics data. The models demonstrated strong performance on clean data, achieving macro F1-scores between 0.90 and 0.93. Even with threat-matched training, performance remained robust, with F1-scores between 0.904 and 0.916. The study also developed a modular framework to evaluate model robustness against realistic telematics data degradation, simulating five corruption families at various severity levels. A key finding was a distinct failure mode: while event-level corruptions (like burst loss or spurious toggles) had only a minor impact on performance, temporal jitter in continuous channels severely degraded macro F1-scores to 0.44-0.50 across all three model types. This highlights a critical vulnerability for these models in real-world scenarios where precise timing might be affected.