Vehicle theft detection by generative adversarial networks on driving behavior

Abstract

Human driving behavior can be a unique fingerprint to identify individual drivers and can be used for vehicle theft detection. Prior research often uses supervised learning to classify drivers’ behaviors. However, that method is not suitable for vehicle theft detection because it is not possible to derive a training dataset which comprehends all possible thieves. A few studies have instead leveraged unsupervised learning, such as k-means, but such approaches cannot achieve acceptable accuracy or provide a method of tolerating the false outcomes from classifying individual data points. Moreover, the influence of environment and road conditions on the driver classification was not included further limiting its usability. In this work, we propose a redesigned Generative Adversarial Network (GAN) model, Convolutional Long short-term GAN (CLGAN), which comprises long short-term memory (LSTM) and a convolutional neural network (CNN). This model ensures the ability of feature extraction from the convolutional layer and feature preservation from the LSTM layer, while simultaneously minimizing overfitting. It is instructive to analyze and benchmark CLGAN with the various GAN models, such as DCGAN, RNN-GAN, and AAE. Applying a public dataset which was generated by an electronic control unit (ECU) and collected through in-vehicle Controller Area Network (CAN) bus, our model achieves high accuracy of 98.5%, and is more robust against various driving conditions on multiple types of roads. In addition, we leverage threshold random walk to ensure the reliability of detection as well as adopting Principal Component Analysis (PCA) on feature pre-processing to improve the accuracy by approximately 20%. This work sheds light on a feasible and practical way of implementing vehicle theft detection.