An Improved Machine Learning Framework for Phishing URL Detection Using the PhiUSIIL Dataset

Abstract

One of the biggest threats to cybersecurity continues to be phishing attacks that involve malicious URLs that attempt to obtain personal information from users. This paper proposes a new method of detecting phishing URLs using the PhiUSIIL dataset, which contains a large number of features that allow for the analysis of the effectiveness of multiple classification algorithms. The new detection methodology includes a preprocessing of the data, feature selection for the classification model via an ANOVA based methodology, and four different supervised classification algorithms: Random Forest, XGBoost, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). In addition, this paper presents multiple performance metrics (i.e., accuracy, precision, recall, and F1-Score) that were used to evaluate the performance of the phishing URL detection models. According to the experiments, all the ensemble models did an outstanding job, with Random Forest and XGBoost having the highest accuracy of 100%, and all other models having more than 99.9% for each evaluation metric compared, indicating that this new framework will model the complexity of features in phishing domains successfully and provide fast computational times. Furthermore, this research further confirms the advantages of using ML and ensemble methods as scalable, dependable, and accurate solutions for detecting phishing URLs in real-world applications.