OFSPUD: Optimizing Feature Selection for Phishing URL Detection through Statistical Tests and Ensemble Methods

The surge in online users utilizing cloud-based platforms—particularly for financial and retail services—is fueled by the convenience and flexibility these services provide. To effectively counter evolving URL phishing threats, machine learning classifiers should analyze all components of URLs, including domain structures, path patterns, and query parameters, to enhance threat detection. The proposed framework, OFSPUD (Optimizing Feature Selection for Phishing URL Detection through Statistical Tests and Ensemble Methods), was evaluated using a large, imbalanced dataset containing 11,055 benign and phishing URLs with over 31 features. Using ANOVA (Analysis of Variance), 25 of the 31 features were selected, significantly improving the method’s performance and correctness. The RBWSV (Rank-Based Weighted Soft Voting) en semble method outperformed prominent machine learning techniques, achieving test accuracy of 97.47% and F1 and validation accuracies of 99% after applying ANOVA based feature selection. Thus, the proposed method, combining ANOVA-based feature selection with the RBWSV ensemble approach, achieved near-perfect accuracy while improving efficiency.