Robust Explainable and Calibrated Machine Learning for Phishing URL Detection under Constrained Feature Perturbations

Phishing detectors can achieve high benchmark accuracy while remaining vulnerable to feature manipulation and difficult to assess as trustworthy decision systems; this study evaluates leakage controlled discrimination calibration explanation stability and constrained adversarial robustness on the PhiUSIIL dataset through a normalized domain grouped development protocol and one frozen test assessment; five model families were compared under conservative and full sensitivity feature conditions while calibration permutation importance Tree SHAP explanations and greedy feature space evasion were analysed; conservative uncalibrated HistGradientBoosting achieved locked test recall of 0.999897 F1 of 0.999949 and Brier score of 0.000041 although attack success reached 0.7042 at three changes and 0.8462 at five changes while mean SHAP cosine similarity fell to minus 0.350; the study also presents a six layer operational architecture that connects evidence governance prediction explanation robustness controls and monitoring; the evidence shows that clean grouped performance and probability reliability can coexist with severe representation level vulnerability and therefore phishing model evaluation should report predictive calibration explanation architectural and adversarial evidence together.