Real-Time Phishing Detection Using Machine Learning on URL Features

Phishing remains one of the most pervasive cybersecurity threats, exploiting users through deceptive URLs to harvest sensitive data. This paper presents a real-time, machine learning-based framework for phishing URL detection that relies solely on lexical and domain-based features, eliminating the need for webpage content or third-party data. A curated dataset of over 11,000 URLs-balanced using SMOTE and preprocessed for consistency-serves as the foundation for training. Feature selection via Canopy Clustering and Recursive Feature Elimination reduced dimensionality to 14 key attributes. A hybrid ensemble model combining Logistic Regression, Support Vector Machine, and Decision Tree using soft voting was developed and outperformed standalone models, achieving 97.3% accuracy, 97.1 % precision, and a 98.9 % true positive rate. The proposed approach is scalable, lightweight, and suitable for deployment in web browsers, email clients, and security gateways. Limitations such as language dependency and content obfuscation are addressed with recommendations for integrating deep learning and multilingual datasets in future work.