An Improved Webpage Phishing Detection Based on an Ensemble Feature Selection and Multilevel Techniques

Over the decades, technology has presented human facets with diverse means of accomplishing complex tasks, especially in the communication domain. One of the risks identified over the decade terrifying network devices and end-users is phishing attacks. Generally, the anti-phishing agents are categorized into conventional and machine learning, but the conventional methods are outdated due to the dynamic nature of the attacks. Furthermore, the machine learning approaches are recognized to be the right antidote to curb the menace, but they involve various steps. One crucial step among the various steps lackadaisically ignored is feature selection and the models are challenged with diverse issues such as overfitting, bias, high computation, and decreased interpretability. To overcome this challenge, this study explored the filter measure methods by applying it as an ensemble feature selection framework to address the complicated issue of feature superfluous, and multilevel models to improve the poor performance of the classical models in the existing studies. The feature obtained by the proposed feature selection and the individual features generated by the filter measures were used. The outcome showed that the features were drastically reduced by the proposed feature selection framework compared to the individual-generated features. However, the models were trained and tested on these four sets of URL features obtained. The results showed that the proposed ensemble feature selection and multilevel mode effectively improved the poor performances of the considered classical models having recorded an accuracy of 98.4% using 7 of the features.