Bachelor of Science in Information Technology

Abstract

Phishing attacks continue to pose a serious cybersecurity threat by imitating trusted websites to deceive users into revealing sensitive information. Traditional blacklist-based solutions and static machine learning models, especially those relying only on URL analysis, are ineffective against newly created phishing websites. Attackers frequently change domains, page layouts, and visual elements, allowing phishing pages to bypass conventional detection techniques and increasing the risk of financial loss, data breaches, and reduced user trust. To overcome these limitations, this project proposes a multimodal phishing detection system that simultaneously analyzes URL features, Document Object Model (DOM) structures, and visual webpage screenshots. The system employs LightGBM for URL and DOM-based features and a Convolutional Neural Network (CNN) for visual feature extraction. A dynamic feature weighting mechanism adaptively balances the contribution of each modality, ensuring robust detection even when attackers attempt to evade individual feature checks. To improve transparency and user trust, the system integrates Explainable AI (XAI) techniques. SHAP is used to highlight influential features, while a language model generates clear, human-readable explanations describing suspicious elements such as deceptive domains, unsafe form actions, or cloned layouts. The solution is deployed as a real-time web application using a React frontend, Node.js backend, and a Flask-based machine learning API, providing an adaptive, interpretable, and practical phishing detection system for real-world use.