A Method for Collecting Vehicular Network Data for Enhanced Anomaly Detection

Advances in anomaly and intrusion detection systems (IDS) for vehicle networks can enable attacks to be stealthier, such as mimicking the behaviours of aggressive driving to disrupt normal vehicle operations. Moreover, data collection, classification and evaluation has been challenging due to vehicle manufacturers deploying proprietary encoding schemes for sensor communications. This paper proposes a method for vehicle network data collection and processing to provide naturalistic datasets for identifying different driving behaviours. Firstly, we provide a guide for the assembly of an Arduino Uno R3 with a Sparkfun CAN bus shield for connection to the on-board diagnostics (OBD) port to collect the vehicle data. Secondly, we offer baseline metrics for labelling passive and aggressive driving behaviours through the analysis of a vehicle's CAN bus data. Finally, a state-of-the-art deep learning model that combines the convolutional neural network (CNN) and recurrent neural network (RNN) architectures is implemented to evaluate the fitness of the data collected using the proposed method. The collected naturalistic dataset was used to train this model in-order to recognise driver behaviour patterns that are otherwise unobservable to humans and that can be used to identify passive and aggressive driver events. Our experiment shows a 0.98 F1 score, indicating our data collection method and metrics used for selecting features can provide quality datasets for intrusion detection against attacks mimicking aggressive driving.