Toward trustworthy cyber defense: A cognitive intrusion detection system with explainability and probabilistic uncertainty calibration

Intrusion Detection Systems (IDSs) are essential for safeguarding digital infrastructure, yet many modern machine learning-based IDSs function as opaque “black boxes” and become unstable under distributional shifts. This paper proposes a Cognitive Intrusion Detection System that integrates explainability, probabilistic calibration, and uncertainty estimation to enhance reliability and transparency in cyber defense. A Random Forest classifier was optimized and calibrated using Platt scaling to improve probabilistic reliability, while explainability was achieved through global and local SHAP analyses. Uncertainty was quantified using a probabilistic entropy-based metric and evaluated for stability across random seeds. Simulations on the CICIDS2017 Friday-Phishing dataset demonstrated near-perfect predictive performance, achieving an average F1-score of 0.9999 ± 0.0001, AUC = 1.0, and a 45% reduction in Expected Calibration Error following calibration. These results confirm that the proposed framework is not only highly accurate but also interpretable, stable, and reproducible. The integration of calibrated probability, explainable reasoning, and uncertainty awareness establishes a trustworthy cognitive foundation for next-generation cyber-defense systems.