The vast usage and development of web technology generate numerous types of web pages. Besides, not all these types are legitimate webpages. Phishing sites mislead web page users into taking harmful actions. However, there is a need for a tool to address this type of problem. Deep learning models are used in dealing with web technology to detect whether the webpage is either legitimate or phishing. Herein, a novel Tabular Multi-Head Attention (TabMHA) model is presented to perform a binary classification task. The main task is to classify whether the webpages are phishing or not. The proposed model is trained and tested on a benchmark dataset related to phishing detection. It contains 5000 legitimate web pages and 5000 phishing ones; the overall is 10,000. Also, the feature numbers in the dataset are out of 48 features. As a consequence, the proposed model achieved a powerful performance compared with other models in the literature; the model achieved an accuracy level of 99.6%. This result is considered a promising result and can be integrated into real-world detection models.

Phishing websites are characterized by distinguished visual, address, domain, and embedded features, which identify and defend such threats. Yet, phishing website detection is challenged by overlapping these features with legitimate websites’ features. As the inter-class variance between legitimate and phishing websites becomes low, commonly utilized machine learning algorithms suffer from low performance in overlapping feature cases. Alternatively, ensemble learning that combines multiple predictions intending to address low inter-class variations in the classified data improves the performance in such cases. Ensemble learning utilizes multiple classifiers of similar or different types with multiple deviations of the training data. This paper develops a framework based on random forest ensemble techniques. The limitations of the random forest are the inability to capture the high correlation between features and their join dependency on the label. The random forest is combined with k-means clustering to capture the feature correlation. The framework is evaluated for phishing detection with a dataset of 5000 samples. The results showed the proposed framework over-performed the random forest classifier, all other ensemble classifiers, and the conventional classification algorithms. The proposed framework achieved an accuracy of 98.64%, precision of 0.986, recall of 0.987, and F-measure of 0.986.