Robustness and Explainability Analysis of Machine Learning Models for Phishing URL Detection

Machine learning models have shown high accuracy for phishing URL detection, yet their robustness under distribution shifts, cross-dataset generalization, and interpretability remain underexplored in practical cybersecurity settings. The study evaluates multiple supervised learning classifiers through testing on various phishing URL datasets which exist in public domain and through their assessment of adversarial perturbation scenarios. Our research evaluates how model performance decreases when we test on new datasets which differ from the original training data while we develop quantitative robustness metrics to measure adversarial resistance. The implementation of explainable AI techniques such as SHAP enables us to interpret model decisions while we analyze how dataset and attack variations affect feature importance. Our results demonstrate that performance and interpretability of systems show considerable changes when tested across different datasets and facing various adversarial attacks which creates a need for development of robust design frameworks to protect phishing detection systems. We propose evaluation guidelines for future research while we show how explainability increases trust and operational understanding without decreasing detection performance.