Emergency Evaluation in Connected and Automated Vehicles

An intelligent transportation system (ITS) provides improved transport efficiency and safety based on vehicle communication. Connected and automated vehicles (CAVs) as part of an ITS are projected to revolutionize the transportation industry, primarily by allowing real-time and seamless information exchange between vehicles and roadside infrastructure. Although these CAVs are expected to offer vast benefits, new problems in terms of safety, security, and privacy will also emerge. Since CAVs continue to rely heavily on vehicle sensors and information obtained from other vehicles and roadside units, abnormal sensors and malicious cyber attacks can lead to destructive results and fatal crashes. Therefore, ensuring reliable and secure information dissemination across vehicles and roadside units is vital for many applications and in the safety-critical aspect of CAVs. As a result, mechanisms that can detect anomalies and identify attack sources in real- time are necessary before the mass deployment of CAVs. This dissertation designs an approach for anomaly detection by utilizing deep Learning (DL), and machine learning (ML) mechanisms, namely Bayesian deep learning (BDL) empowered with discrete wavelet transform (DWT), to detect and identify abnormal behavior in CAVs. The proposed approach’s numerical experiment shows high performance in detecting anomalies and identifying their scores with high accuracy, sensitivity, precision, and F1 - score. Furthermore, this proposed method outperforms baseline BDL and convolutional neural network (CNN) approaches in detecting and identifying anomalies. Performance-wise, the proposed approach is evaluated in terms of the following performance metrics: sensitivity, precision, and F1 - score. Based on the simulation, the proposed approach achieves performance gains of 6.98 %, 9.10 %, and 7.37% over CNN and 11.89 %, 7.32 %, and 9.37% over BDL at duration d = 3 and linspace(0, 6000) for the difficult gradual drift anomaly. In another work, a new architecture of ML-Based Trust (MLBT) mechanism in detecting adversary behaviors in a vehicular-based M2M-C (VBM2M-C) framework is proposed. A combination of extreme Gradient Boost (XGBoost) and binary particle swarm optimization (BPSO) is introduced to detect and identify malicious behaviors within the network. The proposed MLBT is evaluated over different probabilities of attacks. The results of this evaluation show that the proposed approach outperforms the state-of-the-art mechanisms by 10% inaccuracy, 9% in true positive rate (tpr), and lowers false positive rate (fpr) by 9 %, 10% in precision, 8.10% in recall, 9.3% in sensitivity, and 10% in F1 - score with reference to the attacker density of 30% in the selected metrics better than the compared approaches. Moreover, an innovative data-driven approach was equally developed, which involves the combination of discrete wavelet transform (DWT) and double deep Q network (DDQN) method for anomaly detection in CAVs. The DDQN is modified to accommodate classification by taking the state’s data feature while labeling as the action. The features in DWT and DDQN are combined to enhance anomaly detection performance in CAV networks. The DWT smoothens the basic safety messages (BSMs) sensor reading before the BSMs are fed into the DDQN approach. F1 - score and sensitivity are used to access the performance of the proposed method. Overall, the proposed method achieves a performance gain of 20% and 10% at a small density of anomaly distribution and 12% and 8% at a high density of anomaly distribution for ensemble multilayer perceptron (EMLP) and support vector machine (SVM)