Domain-Holdout Evaluation of Machine-Learning Models for Phishing URL Classification

Phishing URLs are used to direct users to deceptive web resources that may attempt to obtain credentials, payment information, or other sensitive data. This paper reports a comparative experiment using the PhiUSIIL Phishing URL dataset and three classifiers: Logistic Regression, Random Forest, and a Multilayer Perceptron (MLP) neural network. The reported methodology uses eleven URL-related features and a registered-domain-held-out evaluation intended to reduce optimistic estimates caused by URLs from the same domain appearing in both training and testing data. In the supplied domain-held-out results file, MLP achieved the highest accuracy (99.5861%), F1-score (99.5087%), and ROC-AUC (0.99837), while Logistic Regression achieved the highest phishing precision (99.9640%) and Random Forest achieved the highest phishing recall (99.1983%). These are results from the supplied experiment output and should not be interpreted as proof of performance on all future phishing campaigns. A split-audit file supplied separately reports a different train/test row count, so the final split metadata and confusion matrix must be regenerated from the same run before public release. Keywords—phishing URL detection, domain holdout, machine learning, Logistic Regression, Random Forest, multilayer perceptron, cybersecurity.