Automobile Theft Detection by Driving Behavior Identification Using Deep Autoencoder

Abstract

Modern vehicles consist of an on-board detection unit that can record a driver’s driving behavior. Detecting anomaly in the driving behavior can be used for theft detection. There are many supervised learning models to detect driving behavior. However, it is impractical to collect the behavior of possible thieves beforehand for training. In this work, we design an unsupervised deep autoencoder model, which can learn the driving behavior only from one or few vehicle owners and it can recognize non-owner driving behavior as the vehicle theft. The model is lightweight and it can achieve high accuracy up to around 98% like a supervised model. The analysis results also show that each driver has different important features for the detection when compared with the other drivers.