An Optimized Machine Learning Framework for Phishing Website Detection Integrating Feature Robustness and Adversarial Resilience Ranking

Phishing attacks continue to be a significant cybersecurity threat, requiring accurate and robust detection mechanisms. This study presents a machine learning-based phishing website detection framework incorporating novel multi-criteria ranking techniques to identify the most effective classifiers. A dataset containing 48 extracted features from phishing and legitimate websites was used to train seven machine learning models, including Decision Trees, k-Nearest Neighbors, SVM, Naïve Bayes, Gradient Boosting, LightGBM, and CatBoost. To enhance model selection, three ranking techniques were applied: Feature Robustness Ranking, which evaluates model stability when features are removed; Generalization Ranking, which assesses performance on unseen phishing attacks; and Adversarial Resilience Ranking, which measures robustness against small manipulations in phishing features. The top three ranked models—CatBoost, LightGBM, and Gradient Boosting—were combined using a Voting Classifier, achieving an accuracy of 97.3%, surpassing individual models. The results demonstrate that the proposed ranking and ensemble-based approach improves phishing detection by ensuring stability, adaptability, and resistance to adversarial modifications. This methodology provides a reliable framework for phishing website classification, contributing to enhanced cybersecurity defenses.