An Integrated Visual–Textual Approach for Accurate Phishing Detection on Social Media Platforms

Abstract

The rapid rise of phishing attacks on social media platforms highlights an urgent and growing cybersecurity challenge in today’s digitally connected world. As millions of users rely on these platforms for communication, business, and information sharing, cybercriminals increasingly exploit them to distribute deceptive links, fake advertisements, and fraudulent messages designed to steal sensitive data such as login credentials, financial information, and personal identities. The sophistication of these attacks—often leveraging social engineering techniques, impersonation of trusted entities, and AI-generated content—makes them difficult for users to recognize and avoid. Consequently, there is a critical need for robust and intelligent detection systems capable of reliably identifying malicious content in real time. Such systems should incorporate advanced machine learning models, behavioral analysis, and continuous adaptation to emerging attack patterns in order to effectively mitigate risks, protect users, and maintain trust in social media ecosystems. This paper proposes a multimodal phishing detection framework that integrates visual indicators (e.g., logos and page imagery) with textual signals (e.g., URL structure and phishing-related keywords). Our contributions are threefold: (i) a feature fusion strategy that combines visual and textual cues, (ii) the construction of a comprehensive dataset using feature extraction and natural language processing (NLP), and (iii) a comparative evaluation of multiple machine learning models. We extract 27 discriminative URL features and evaluate Random Forest, XGBoost, LightGBM, and Naı̈ve Bayes. Random Forest achieves the best performance, reaching 98.6% accuracy with a false positive rate of 0.01, demonstrating the effectiveness of the proposed multimodal approach for phishing detection on social media.