Phishing Website Detection: A Comparative Analysis of Outlier Detection, Feature Selection, and Ensemble Algorithms

Abstract

With the advancement of network technologies, phishing is a social engineering attack where attackers use fake websites to steal users' identities, and particularly their financial information. Machine learning methods have been widely used in recent years to detect these attacks. In this study, gradient-based ensemble models (GBM, XGB, LGBM, HGBM) were used for phishing detection. First, anomaly detection was performed on the original dataset using the iForest and IQR algorithms; then, sub-datasets named IFAS, IMIS, IQAS, and IQMIS were created using ANOVA and Mutual Information (MI) methods. Following hyperparameter optimization with Optuna, the XGB model achieved 95.67% and 92.87% accuracy on the IMIS and IQMIS datasets, respectively. The LGBM model stood out with a 95.53% accuracy rate on the IFAS dataset, while the highest accuracy on the IQAS dataset was achieved by the HGBM model at 93.66%. The results demonstrate that the XGB model has high potential for success in phishing detection.