An Intelligent TabNet-based Framework for Phishing Website Detection using URL and Webpage Features

Abstract

One of the most common cyber security threats today is through phishing websites, which appear legitimate to trick users into providing their details, including login information, banking details, and personal information. The sophistication of phishing attacks has grown and so has the application of machine learning and deep learning algorithm techniques to intelligent and automated phishing website detection. Traditional machine-learning methods, however, typically require a lot of features and are not able to capture the complex relationships between high-dimensional tabular features. This has generated a growing interest in creating deep learning models that are tailored to structured data. In this work, a phishing website detection framework based on TabNet is proposed, which uses an automatic feature selection and sequential attention mechanism to increase the classification performance. The proposed model is tested with the Phishing URL Dataset from PhiUSIIL which contains 235,795 website instances, 55 features extracted from the URL and four classes. The model includes data preprocessing, embedding of features, TabNet encoder consisting of Feature Transformer and Attentive Transformer modules, and a SoftMax classifier for final prediction. The experimental results show that the proposed model has good accuracy, precision, recall, and F1 score of 98.0%, 98.0%, 98.0%, and 98.0%, respectively, which reflects its strong and stable classification performance. The results confirm the effectiveness of the proposed TabNet framework for improving phishing website detection and its potential applicability in real-world cybersecurity systems like secure web browsing, enterprise security, and financial fraud prevention.