Recent developments in intelligence technologies have led to an explosion in the use of connected and automated vehicles (CAVs). Unfortunately, these autonomous vehicles face an increasing risk of vulnerability due to various attacks. Vehicle intrusion detection mechanisms are widely employed to mitigate the threats. Although some works have addressed this issue, few works consider situations that the attack detection may be fooled via simultaneous attacking multiple sensors. In this paper, we propose a novel intrusion detection system integrated Space Dimension Model and Time Dimension Model based on sensor data fusion for countering both independent attack and confederate attack in automated vehicles. In Space Dimension Model, the correlations of multivariate in-vehicle sensor data among multiple sensors are utilized as the input of our optimized convolutional neural network (CNN) model. Especially, we construct vehicle state matrices to characterize the underlying correlation of data between each sensor and other sensors, and then input them into our network for classification. To describe the abrupt deviation caused by anomalous multivariate sensor data itself, we design Time Dimension Model to capture the sensor behaviors of vehicle state vectors at adjacent time by utilizing the Mahalanobis distance (MD) metric. Extensive empirical studies based on a real-world vehicular dataset demonstrate the effectiveness of our integrated anomaly detection mechanisms by a comparative analysis with two detection models that consider only Space Dimension and Time Dimension respectively. The method can outperform the related work under different scenarios 1) with the gain of up to 3.01% in accuracy and 3.04% in F1 score; 2) with ability of defending against confederate attack effectively.

Connected and automated vehicles (CAVs) involving massive advanced sensors and electronic control units (ECUs) bring intelligentization to the transportation system and conveniences to human mobility. Unfortunately, these automated vehicles face security threats due to complexity and connectivity. Especially, the existing in-vehicle network protocols (e.g., controller area network) lack security consideration, which is vulnerable to malicious attacks and puts people at large-scale severe risks. In this paper, we propose a novel anomaly detection model that integrates a continuous wavelet transform (CWT) and convolutional neural network (CNN) for an in-vehicle network. By transforming in-vehicle sensor signals in different segments, we adopt CWT to calculate wavelet coefficients for vehicle state image construction so that the model exploits both the time and frequency domain characteristics of the raw data, which can demonstrate more hidden patterns of vehicle events and improve the accuracy of the follow-up detection process. Our model constructs a two-dimensional continuous wavelet transform scalogram (CWTS) and utilizes it as an input into our optimized CNN. The proposed model is able to provide local transient characteristics of the signals so that it can detect anomaly deviations caused by malicious behaviors, and the model is effective for coping with various vehicle anomalies. The experiments show the superior performance of our proposed model under different anomaly scenarios. Compared with related works, the average accuracy and F1 score are improved by 2.51% and 2.46%.