Impact Analysis of Filter and Wrapper-Based Feature Selection Techniques for Webpages Phishing Attacks Identification

Abstract

Phishing, which involves fraudulently gaining access to sensitive assets of unsuspecting individuals through deceptive and malicious emails, is a major global threat to internet users. The proliferation of phishing sites and their operations is occurring at an alarming rate, raising significant concerns about how to forestall them. Numerous research efforts are underway to detect phishing attempts before they can compromise important information and cause damage. Compared to conventional methods, machine learning has proven highly effective at detecting phishing attacks by analyzing different features. This study analyzed the behaviors of seven classification data mining algorithms on optimal subset features selection using Wrapper (Boruta) and Filter-based (Mutual-Information). Real-life phishing webpage datasets were used for the analysis. Ensemble classifiers such as Voting, Gradient Boosting, and Random Forest were used in the experiments. Two experiments were conducted. In the first experiment, K-Nearest Neighbor (K-NN) yielded the highest accuracy among single classifiers, with a score of 94.1%, while Random Forest (RF) ensemble achieved 96.7%. In the second experiment, using another baseline feature set, RF performed excellently under the Boruta method with an accuracy of 97.25%, while K-NN retained the highest accuracy of 95.20% among single classifiers. This study provides empirical evidence that feature selection techniques have a great impact on the performance of ML models, for both single and ensemble classifiers, in the detection of phishing attacks.