Vision GNN Based Phishing Website Detection

Phishing, which is a common cyber threat, exploits human vulnerabilities by impersonating legitimate entities to deceive users into sharing sensitive information. Traditional methods of detecting phishing websites primarily rely on textual features and heuristics, often failing to capture the evolving sophistication of phishing tactics. This study introduces an innovative approach for identifying phishing websites through an image of the website utilizing VisionGNN architecture which is a Graph Neural Networks (GNN) based approach. By using the VisualPhish dataset, consisting 1195 pages across 155 websites, our approach focuses on significantly enhancing accuracy in phishing website detection using RGB images of the websites. Through training with image augmentations and the application of optimization techniques such as AdamW and cosine decay for learning rate adaptation, a good result in predictive capabilities are attained. The outcomes demonstrate the model’s efficacy in accurately classifying both benign and phishing websites. Impressively, our VIsionGNN-based approach with its pyramidsmall model architecture out of its four combinations achieves 97% accuracy on the VisualPhish data set. With our approach being the first time VisionGNN architecture has been utilized in phishing website detection, our study spotlights the potential of pioneering Graph Neural Network-based approaches in strengthening cyber security against the escalating threat of phishing websites. Furthermore, this study lays the groundwork for future research areas aimed at utilizing VisionGNN-based approaches for more real-world applications.