Explainable Hybrid Ensemble Learning for Deceptive Website Detection with Cross-Dataset Validation

Abstract

Phishing websites are becoming more adept at using phony websites and similar URLs to fool consumers into believing they are legitimate to obtain important sensitive data, including credit card numbers and passwords. Blacklist-based techniques are unable to validate dynamic deceptive websites and freshly registered domains, showing that machine learning algorithms need to provide a more robust and reliable result. This work presents an explainable framework that benchmarks conventional machine learning models and gradient boosting methods before combining CatBoost, XGBoost, and LightGBM into a hybrid stacking ensemble for deceptive website detection. 30 aligned features were used across two datasets, Kaggle and UCI, after a uniform preprocessing and feature-alignment pipeline. Boosting-based models consistently outperformed conventional classifiers, and the tuned hybrid stacking ensemble was competitive with the strongest individual boosting models, tying for the top accuracy on Kaggle (97.38%) and closely trailing the best individual model on UCI (97.83% vs. 97.87%), while reaching the single highest cross-dataset accuracy of 98.67% when trained on Kaggle and tested on UCI. This cross-dataset result is an encouraging indicator of generalization beyond a single data distribution, though it reflects evaluation on two public datasets rather than newly emerging or temporally separated phishing samples. The framework's explainability through SHAP, which improves transparency and reliability by highlighting pertinent URL and domain properties relevant to phishing detection, makes it dependable and useful for real-world deployment.