Temporal learning techniques for vehicle cybersecurity

Driver identification is being used lately for vehicle anti-theft and identifying fake driver accounts based on their driving behaviors. Anti-theft is a challenging problem in vehicle industries, as it depends on external devices to protect against vehicle burglary. The embedded sensors in the vehicles are indispensable for modern control systems for enhancing driving safety. The sensing data was originally used for maintenance. We can extract such data through the OBD (On-Board Diagnostics) port and utilize them. Recently, researchers began to use sensing data extracted from car sensors for the driver classification problem and employing the highest similarity for driver verification. However, they encountered several hindrances in producing a stable driver identification model for addressing the cold start issue and long patterns. Furthermore, some approaches struggled with the unpleasant performance of the generated models when they increased the action space (> 2 drivers). This dissertation aimed to enhance multivariate time series (MTS) analytics for driver identification. In this dissertation, we proposed two novel approaches called LiveDI (Live Driver Identification) and OnlineDC (Online Driver Classification), which leverages temporal driving behaviors in order to identify the driver behind the wheel. In this study, we also considered protecting the human subjects’ privacy. LiveDI and OnlineDC provide two preprocessing algorithms for generating features from the driving behaviors (MTS). The two feature generation algorithms are the segmented feature generation algorithm (SFG) and the spectral/ temporal/ statistical feature generation algorithm (FG). SFG computes 11 statistical features on each segment per signal of the sensing data. SFG helped in decreasing the state space and declining the training time. It also enhances the performance of the driver identification model. While, FG generates 82 valuable features (the spectral, temporal, and statistical features) from each sequence. FG unifies the input size for different lengths of MTS, and it generates several kinds of features which help in producing a scalable and stable model. It also enhanced the classification model in addressing the cold start problem. Additionally, the spectral features helped in reducing the signal noise. We had designed two deep neural networks to improve and stabilize the performance of the classification models. MGRU-FCN combines the MTS Gated Recurrent Unit (GRU) and the Fully Convolutional Networks (FCN). Also, we presented MGRU-ResSE (MTS GRU with Residual Neural Network with Squeeze-and-Excite block ”ResSE”). We compared our deep neural networks with traditional classification algorithms and other neural networks to show the importance of these new networks in enhancing and stabilizing the performance of the driver identification models.