Evaluating machine learning-based intrusion detection systems with explainable AI: enhancing transparency and interpretability

Abstract

Machine Learning (ML)-based Intrusion Detection Systems (IDS) are integral to securing modern IoT networks but often suffer from a lack of transparency, functioning as “black boxes” with opaque decision-making processes. This study enhances IDS by integrating Explainable Artificial Intelligence (XAI), improving interpretability and trustworthiness while maintaining high predictive performance. Using the UNSW-NB15 dataset, comprising over 2.5 million records and nine diverse attack types, we developed and evaluated multiple ML models, including Decision Trees, Multilayer Perceptron (MLP), XGBoost, Random Forest, CatBoost, Logistic Regression, and Gaussian Naive Bayes. By incorporating XAI techniques such as LIME, SHAP, and ELI5, we demonstrated that XAI-enhanced models provide actionable insights into feature importance and decision processes. The experimental results revealed that XGBoost and CatBoost achieved the highest accuracy of 87%, with a false positive rate of 0.07 and a false negative rate of 0.12. These models stood out for their superior performance and interpretability, highlighting key features such as Source-to-Destination Time-to-Live (sttl) and Destination Service Count (ct_srv_dst) as critical indicators of malicious activity. The study also underscores the methodological and empirical contributions of integrating XAI techniques with ML models, offering a balanced approach between accuracy and transparency. From a practical standpoint, this research equips human analysts with tools to better understand and trust IDS predictions, facilitating quicker responses to security threats. Compared to existing studies, this work bridges the gap between high-performing ML models and their real-world applicability by focusing on explainability. Future research directions include applying the proposed methodology to more complex datasets and exploring advancements in XAI techniques for broader cybersecurity challenges.