EVIDENCE-BASED GOVERNANCE FOR TRUSTWORTHY AI: EXPLAINABLE ENSEMBLE INTRUSION DETECTION BENCHMARKS FOR CRITICAL INFRASTRUCTURE

The proliferation of Artificial Intelligence (AI) in U.S. critical infrastructure introduces unprecedented operational capabilities alongside complex security, privacy, and trust challenges. While theoretical frameworks dominate existing literature, empirical validation remains limited, hampering practical implementation and standardized governance. This study addresses these gaps by quantitatively analyzing AI-driven intrusion detection using ensemble machine learning models—Random Forest and Gradient Boosting—applied to the NSL-KDD dataset. Both models achieved exceptional classification performance, with accuracy, precision, recall, F1-score, and AUC all at 100%. Random Forest demonstrated superior generalization, with only 12 misclassifications out of 9,147 test instances (99.87% effective accuracy), achieving false-positive and false-negative rates of 0.13% each. Gradient Boosting showed 18 misclassifications (99.80% effective accuracy), with false-positive and false-negative rates of 0.20% each. Feature importance analysis identified forward packet length mean as the strongest predictor, accounting for 32.4% of Random Forest's decision-making and 82.7% of Gradient Boosting's strategy. SHAPbased interpretability enhances transparency and accountability, critical for stakeholder trust. These findings provide quantifiable benchmarks for policymakers developing standardized AI governance frameworks, though external validation with real-world traffic is necessary to address potential overfitting. This research establishes an empirical foundation for secure-by-design principles and regulatory harmonization across critical infrastructure sectors.