Deep Real-Time Anomaly Detection for Connected Autonomous Vehicles

Abstract Connected and autonomous vehicles (CAV) are expected to change the landscape of the automotive market. They are autonomous decision-making systems that process streams of observations coming from different external and on-board sensors. CAV like any other cyber-physical objects are prone to signal interference, hardware deterioration, software errors, power instability, and cyber-attacks. To avoid these anomalies which can be fatal, it is mandatory to design a robust real-time technique to detect them and identify their sources. In this paper, we propose a deep learning approach which consists of hierarchic models to firstly extract the signal features using an LSTM auto-encoder, then perform an accurate classification of each signal sequence in real-time. In addition, we investigated the impact of the model parameter tuning on the anomaly detection and the advantage of channel boosting through three scenarios. The model achieves an accuracy of 95.5% and precision of 94.2%.