META-ENSEMBLE APPROACH FOR PHISHING WEBSITE DETECTION: COMBINING THE STRENGTHS OF MULTIPLE MACHINE LEARNING MODELS

Abstract

This paper presents a meta-ensemble framework for phishing website detection that combines multiple machine learning models to enhance classification accuracy and robustness. Our approach integrates traditional classifiers such as Random Forest and SVM with advanced models including XGBoost, LightGBM, and CatBoost through voting, stacking, and bagging techniques. Experiments conducted on a comprehensive dataset of phishing and legitimate websites achieved a remarkable accuracy of 97.3% using our meta-ensemble method, outperforming individual models and basic ensembles. Feature importance analysis revealed that SSL certification status, URL characteristics, and domain registration length were among the most significant indicators for phishing detection. The proposed framework demonstrates excellent generalization capabilities while maintaining low false positive rates, making it suitable for real-world cybersecurity applications. This study contributes to the advancement of anti-phishing systems by effectively leveraging the complementary strengths of diverse machine learning algorithms through a hierarchical ensemble architecture.