The advent of autonomous driving systems has given rise to pressing cybersecurity issues regarding the vulnerability of vehicular ad hoc networks (VANETs) to potential attacks. This critical security problem necessitates the application of artificial intelligence (AI) models for anomaly detection in VANETs of autonomous vehicles (AVs). However, the lack of explainability of such AI-based anomaly detection models presents challenges. This motivates an emerging research direction of utilizing explainable AI (XAI) techniques to elucidate the behaviors of anomaly detection models in AV networks. In this work, we propose an end-to-end XAI framework to interpret and visualize the anomaly detection classifications made by AI models securing VANETs. We evaluate the framework on two real-world autonomous driving datasets. The framework furnishes both global and local explanations for the black-box AI models using two XAI methods. Moreover, we introduce two novel feature selection techniques to identify the salient features contributing to anomaly detection, derived from the popular SHAP XAI method and the accuracy of six different black-box AI models.We compare our proposed feature selection approaches with six state-of-the-art feature selection techniques (including two wrapper-based feature selection methods), demonstrating superior performance on various evaluation metrics. To generalize the impact of our feature selection methods, we apply three independent classifiers to evaluate our proposed feature selection approaches. The novel feature selection methods effectively distill themost explanatory features, enhancing model interpretability. Finally, we assess the efficiency (how quickly the XAI models can yield explanatory findings) for each of the six black-box AI models we employed on our two datasets, identifying the most efficient model. By furnishing explanations and visualizations of anomaly detection by AI models, our XAI framework can help in enabling trust and transparency for securing vehicular networks.

The recent advancements in autonomous driving come with the associated cybersecurity issue of compromising networks of autonomous vehicles (AVs), motivating the use of AI models for detecting anomalies on these networks. In this context, the usage of explainable AI (XAI) for explaining the behavior of these anomaly detection AI models is crucial. This work introduces a comprehensive framework to assess black-box XAI techniques for anomaly detection within AVs, facilitating the examination of both global and local XAI methods to elucidate the decisions made by XAI techniques that explain the behavior of AI models classifying anomalous AV behavior. By considering six evaluation metrics (descriptive accuracy, sparsity, stability, efficiency, robustness, and completeness), the framework evaluates two well-known black-box XAI techniques, SHAP and LIME, involving applying XAI techniques to identify primary features crucial for anomaly classification, followed by extensive experiments assessing SHAP and LIME across the six metrics using two prevalent autonomous driving datasets, VeReMi and Sensor. This study advances the deployment of black-box XAI methods for real-world anomaly detection in autonomous driving systems, contributing valuable insights into the strengths and limitations of current black-box XAI methods within this critical domain.

The inherent limitations of individual AI models underscore the need for robust anomaly detection techniques for securing autonomous driving systems. To address these limitations, we propose a comprehensive ensemble learning framework specifically designed for anomaly detection in autonomous driving systems. We comprehensively assess the effectiveness of ensemble learning models for detecting anomalies in autonomous vehicle datasets, focusing primarily on the VeReMi and Sensor datasets. Ensemble techniques are rigorously evaluated against individual models on binary and multiclass classification tasks. The analysis reveals that ensemble models consistently outperform individual models in terms of accuracy, precision, recall, false positive rates, and F1-score. On the VeReMi dataset, ensembles achieve high performance for binary classification, with a maximum accuracy of 0.80 and F1-score of 0.86, surpassing single models. For the Sensor dataset, ensemble models like CatBoost exhibit perfect accuracy, precision, recall, and F1-score, exceeding single models by 11% in accuracy. In VeReMi multiclass classification, Stacking and Blending gave a 5% increase in accuracy compared to single models. Moreover, XGBoost and CatBoost demonstrate perfect recall. Our proposed method enhanced performance despite the increased runtime required by ensemble models. In evaluating false positive rates, ensemble learning demonstrated significant gains, reducing false positives and thereby enhancing overall system reliability.