Context-Aware Anomaly Detection Using Vehicle Dynamics

Replacing traditional vehicular components with electronic components brings numerous benefits but also introduces new vulnerabilities. To cope with this double-edged trend, we propose Context-Aware Detection of abnormal vehicle Dynamics (CADD) in general, or abnormal vehicle accelerations in particular. To account for the limited data availability common in production vehicles, we propose a new detection mechanism based on estimated vehicular contexts, instead of the commonly used “predict-input-then-compare.” That is, without relying on the unrealistically assumed availability of detailed measurements for accurate behavior modeling and prediction, CADD utilizes four sets of vehicle data to perform anomaly detection by cross-validating estimations of the underlying driving contexts, including road inclination, tire slippage, and total mass. Our extensive evaluation with > 87,000 test-cases has shown CADD to achieve > 96% recall and < 0.5% false positive rate. Furthermore, CADD can efficiently pinpoint the anomalous group of data with > 95% accuracy when the vehicle’s behavior deviates 0.07g (0.69 m/s2) from its normal pattern.