A Comparative Study on the Performance of Explainable Artificial Intelligence-Based Intrusion Detection Systems

Abstract

In the modern cybersecurity landscape, the increasing sophistication and diversity of network attacks have significantly reduced the effectiveness of traditional Intrusion Detection Systems (IDS). At the same time, Artificial Intelligence (AI)-based IDS models often operate as black boxes, limiting transparency and hindering their adoption in security-critical environments where explainability and trust are essential.
    To address these challenges, this study proposes an interpretable and efficient IDS framework based on Explainable Artificial Intelligence (XAI). The approach enhances interpretability by reducing feature redundancy and integrating multiple feature selection (FS) methods with machine learning (ML) models to identify the most informative and non-redundant features. In addition, a systematic evaluation framework is introduced to assess the quality, stability, and reliability of explanations generated by XAI techniques. 
    Experimental results demonstrate that combining the minimum Redundancy Maximum Relevance (mRMR) feature selection method with the XGBoost classifier yields superior performance, achieving 99.99% Accuracy. Furthermore, TreeSHAP and Permutation Importance outperform other explanation methods in terms of interpretability, consistency, and stability.
    Overall, this work demonstrates that integrating feature selection techniques with XAI methodologies can substantially improve both detection performance and interpretability in IDS. The proposed framework not only enhances predictive Accuracy but also provides transparent and reliable explanations, thereby increasing trust and supporting practical deployment in real-world cybersecurity systems.