A lightweight intrusion detection system for connected autonomous vehicles based on ECANet and image encoding

Abstract

The Controller Area Network (CAN) bus plays an essential role in Connected Autonomous Vehicles (CAVs), yet its inherent design limitations regarding data protection make it susceptible to malicious intrusions. Contemporary research in intrusion detection predominantly employs Long Short-Term Memory (LSTM) models to analyze CAN IDs as time series data. However, the high computational complexity of LSTM models makes them unsuitable for resource constrained in-vehicle network. To address this problem, a lightweight IDS combining image encoding and an Efficient Channel Attention (ECA) network is proposed. Specifically, three temporal image encoding techniques, Gramian Angular Sum Fields, Markov Transition Fields, and Recurrence Plots are employed to transform CAN ID time-series data into single-channel images, which are then superimposed into three-channel images. A lightweight three-layer convolutional neural network integrated with an ECA module dynamically adjusts channel weights for image classification. Evaluated on real in-vehicle datasets, the method achieves classification accuracies of 99.83%, 99.98%, and 98.75% across three test scenarios with 5.5ms average inference time, demonstrating robust detection capability and computational efficiency.