Trade-off Between Performance, Efficiency, and Explainability in AI-Based Intrusion Detection Systems: A Systematic Review of Models, XAI, and Research Challenges

Intrusion detection system, SHAP, LIME, XAI, Deep Learning, Machine Learning, Post-hoc interpretable.

Authors

June 8, 2026
June 9, 2026

Downloads

Artificial Intelligence (AI)-based Intrusion Detection Systems (IDS) have significantly enhanced the detection of cyber-attacks through the adoption of machine learning and deep learning techniques. However, achieving high detection results often comes at the expense of interpretability, computational efficiency, and generalization. This paper introduces a systematic review of AI-based IDS, focusing on the integration of deep learning algorithms with Explainable Artificial Intelligence (XAI) methods. The paper analyzes a varied range of techniques, including traditional machine learning, deep learning, and hybrid models, alongside commonly used XAI methods such as SHAP and LIME. Furthermore, the review studies widely used benchmark and real-world datasets, emphasize their limitations in representing modern cyber-attack scenarios. Evaluation metrics used in the literature are also critically analyzed, revealing a lack of comprehensive assessment frameworks that incorporate performance, efficiency, and explainability. The findings determine a clear trade-off between accuracy, interpretability, and computational complexity, with most high-performing models lacking intrinsic transparency and relying heavily on post-hoc explanation methods. Finally, this paper recognizes key research gaps, including a limited combination of XAI within model architectures, over-reliance on outdated datasets, and the absence of unified Intrusion Detection Systems frameworks. These insights deliver a foundation for developing next-generation Intrusion Detection Systems solutions that balance performance, explainability, and real-world applicability.