Adaptive Phishing Detection Using Machine Learning: A Novel URL-Based Feature Analysis System

Abstract: Phishing attacks continue to pose a significant cybersecurity challenge worldwide. This study presents a robust, adaptive URL-based phishing detection framework that integrates Balanced Random Forest, and XGBoost within a soft-voting ensemble architecture. The model utilizes lexical, structural, and domain-level features extracted directly from URLs, allowing real-time prediction without relying on blacklists. Experimental evaluation achieved an accuracy of 93.29%, precision of 92.30%, recall of 95.26%, F1-score of 93.76%, and ROC-AUC of 0.982. The ensemble demonstrates strong adaptability in identifying zero-day phishing URLs and can be seamlessly deployed via Flask-based APIs and browser extensions.

Keywords: Phishing Detection, Machine Learning, Ensemble Learning, Cybersecurity, URL Features.