Intrusion Detection for In-Vehicle CAN Bus Based on Lightweight Neural Network

Abstract

With the rapid development of automobile intelligent and networking, substantial information is exchanged between in-vehicle network system and the outside world, thereby threatening the automobile security. Intrusion detection is an important technology to realize the security of in-vehicle networks. The existing research on in-vehicle network intrusion detection mainly focuses on the improvement of detection accuracy, but it lacks consideration of timeliness, whereas the in-vehicle network is a time-sensitive system. This study proposes an anomaly detection method for in-vehicle Controller Area Network (CAN) based on lightweight neural network to reduce the operation time while maintaining the detection accuracy. The redundant neuron screening method and model compression algorithm for layer-by-layer neuron pruning are designed. This presented method can delete the neurons with small contribution and obtain lightweight neural network model. The detection performance of model compression and noncompression is compared through experiments. Results show that under the two real in-vehicle datasets, the detection time is accelerated by 47.7 times and 34.2 times at most, and the average accuracy is increased by 14.5% and 15.7%.