READ: Rule-based Explanation using Analysis of Data to explain the decision of in-vehicle intrusion detection system

Connected Autonomous Vehicles (CAVs) require efficient lightweight machine learning (ML) and deep learning (DL) techniques to secure in-vehicle communication, typically conducted via the Controller Area Network (CAN), which lacks basic security features such as authentication and encryption. Numerous ML/DL-based approaches have been developed to safeguard CAN buses, but their limited decision transparency hinders deployment in critical domains like security. The black-box nature of these models limits their deployment in safety-critical environments due to the lack of transparency and interpretability. Existing explainable artificial intelligence (XAI) approaches, including SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and RuleXAI, provide post-hoc explanations but often generate complex, redundant, and poorly controllable outputs, reducing their practical usability. To address these limitations, we propose READ (Rule-based Explanations using Analysis of Data) , a novel rule-induction-based explainability framework built upon Simple Rule Induction (SRI) . READ comprises two core algorithms: L_READ for local explanations and G_READ for global explanations, enabling both instance-level and model-level analysis. Unlike existing methods, READ provides tunable control over rule complexity , generating compact and human-interpretable explanations while preserving high fidelity to the underlying model. We evaluate READ on Random Forest (RF) and Deep Neural Network (DNN)-based IDS using the CarChallenge2020 dataset . Experimental results demonstrate that for RF-based IDS, READ significantly improves interpretability by reducing the number of rules from 59 to 14 (approximately 76% reduction) and average rule length from 9.25 to 2.4, while achieving high fidelity of up to 94.5%, compared to 70.0% for RuleXAI. For DNN-based IDS, READ reduces rules from 40 to 11 (approximately 72.5% reduction) and achieves fidelity of 95.9%, outperforming baseline methods. Furthermore, READ enables identification of vulnerable features through local explanations, facilitating post-hoc analysis and potential mitigation of incorrect model decisions. These results demonstrate that READ provides a scalable and effective solution for generating compact, controllable, and high-fidelity explanations , improving transparency and trust in ML/DL-based security systems for CAVs.