Optimised Structure of Convolutional Neural Networks for Controller Area Network Classification

Abstract

Security researchers have proved that current modern automobiles are vulnerable to attack, particularly in Controller Area Network (CAN) which controls most of the critical parts in cars. The adversaries can gain access to the compromised vehicle's component and flood with modified CAN packet to cause physical effects. Hence, categorising normal CAN packets had become significant to determine the standard behaviour of CAN bus traffic in detecting attacks. A proposed anomaly detection classifier in this paper is inspired by the sequence classification in Natural Language Processing (NLP) problem, where, the combination of word embedding and Convolutional Neural Network (CNN) algorithm are used. This approach aims to construct a baseline classifier of normal CAN DATA fields according to their CAN ID family. The cross-entropy loss is used to measure the proposed classifier's performance index. Besides, the hyperparameter tuning structure of the classifier is designed based on Taguchi method. The analysis suggested that maximising Signal-to- Noise (S/N) ratio by setting Rectified Linear Unit (Relu) for activation function, epochs of 6, vocab size of 356 and ‘Dropout’ of 0.6, hence prediction loss can be significantly reduced. A systematic analysis design using Taguchi method is considered a new methodology to anomaly detection classifier in CAN bus data.