E-RXAI-IoT: A Systematic Evaluation Framework of Rule-Based XAI Methods for Anomaly Detection in IoT Systems

The increasing complexity of Internet of Things (IoT) systems necessitates the development of interpretable and robust artificial intelligence (AI) techniques for anomaly detection. While prior work has explored deep learning-based explainable AI (XAI) methods, systematic studies evaluating rule-based XAI approaches for IoT anomaly detection remain limited. In this paper, we investigate two state-of-the-art rule-based XAI methods—ANCHOR and RuleFit—across multiple machine learning models: Random Forest (RF), Decision Tree (DT), Deep Neural Network (DNN), and Support Vector Machine (SVM). We introduce a comprehensive evaluation framework that assesses these XAI methods using seven critical metrics mapped to AI and security domains: descriptive accuracy, sparsity, efficiency, stability, fidelity, coverage, and precision. Our framework is applied to two distinct IoT datasets: the N-BaIoT dataset, which captures IoT botnet attacks (Mirai and Gafgyt) on commercial devices, and the MEMS dataset, which focuses on detecting anomalies in sensor readings from manufacturing processes. Through this study, we systematically highlight the strengths and limitations of rule-based XAI techniques in the context of IoT anomaly detection, providing practical insights into their interpretability and robustness. The source code for our evaluation framework is made publicly available for the community to foster reproducibility and encourage further research in this domain with more methods, metrics, and datasets.