A Multi-Modal Phishing Detection System
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Phishing attacks have evolved significantly, employing visual mimicry, semantic deception, and network-level manipulation to bypass traditional detection systems. Conventional approaches based on URL blacklists or single-modal feature analysis often fail against zero-day and dynamically generated phishing pages. This paper presents a multi-modal phishing detection framework that integrates URL lexical features, HTML/DOM structural attributes, visual cues, semantic content, and network-based indicators. Structured features are processed using a stacked ensemble model comprising Logistic Regression, LightGBM, and Linear SVM classifiers. Webpage screenshots are analyzed using a fine-tuned EfficientNet-B0 model to extract visual embeddings, while semantic representations are generated using DeBERTa-v3 Base to identify deceptive language patterns. These heterogeneous features are fused through dense neural layers to produce a final phishing probability score. The system incorporates cost-sensitive learning to address class imbalance and integrates explainability mechanisms, including Grad-CAM visualization and DOM-level feature highlighting. The proposed architecture aims to deliver a scalable, adaptive, and interpretable solution for detecting modern phishing attacks across multiple content modalities.
Fields of Study & Subjects
Spam and Phishing Detection
Misinformation and Its Impacts
User Authentication and Security Systems
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 30 March 2026
- Language
- EN