Hybrid Model for Phishing Website Detection Using Transfer Learning

Authors

• Atul Dubal, Department of Electronics and Computer Science, Pillai HOC College of Engineering and Technology, Rasayani, Khalapur, 410207, Maharashtra, India
• Mansi Subhedar, Department of Electronics and Computer Science, Pillai HOC College of Engineering and Technology, Rasayani, Khalapur, 410207, Maharashtra, IndiaFollow
• Santosh Dhamala, Department of Electronics and Computer Science, Pillai HOC College of Engineering and Technology, Rasayani, Khalapur, 410207, Maharashtra, India
• Manasi Patil, Department of Electronics and Computer Science, Pillai HOC College of Engineering and Technology, Rasayani, Khalapur, 410207, Maharashtra, India

Abstract

The rapid digitization of human activities has intensified reliance on internet-based platforms, creating fertile ground for cybercriminal exploits such as phishing. Despite advancements in detection mechanisms, phishing attacks continue to evolve, leveraging sophisticated visual mimicry to deceive users. This paper proposes a robust vision-based phishing detection system using ensemble deep learning to analyse webpage screenshots. The framework integrates transfer learning with pre-trained VGG16 and DenseNet121 models, extracting complementary low-level texture features (edges, gradients) and high-level hierarchical patterns (logos, layouts). These features are fused through a custom classifier with dropout regularization to mitigate overfitting. A balanced dataset of 3,000 webpage screenshots is curated augmented with horizontal flips, shear, and brightness adjustments to enhance generalizability. 5-fold cross validation techniques are used, where dataset is divided into 5 equal parts, where each one-fold served once as the test set while the remaining four used for training. The model achieves a training accuracy of 93.85% and testing accuracy of 89.00%, with a ROC-AUC score of 95.12% demonstrating strong separability between classes. Key contributions include a novel fusion of VGG16 and DenseNet121 for phishing detection, a lightweight architecture optimized for real-time inference (0.3 seconds per image), and a publicly available dataset to foster reproducibility. Experimental results surpass traditional HOG (Histogram of oriented gradients) based methods by 15% in F1-score and single-model baselines by 6% in accuracy. This work underscores the efficacy of ensemble deep learning in combating visually deceptive phishing attacks, offering a deployable solution for browser plugins or email security systems.

Keywords

Phishing detection, deep learning, transfer learning, VGG16, DenseNet121, ensemble models, computer vision, cyber security

Publication Date

2026

Received Date

04/06/2025

Revised Date

20/02/2026

Accepted Date

01/03/2026

Recommended Citation

Dubal, Atul; Subhedar, Mansi; Dhamala, Santosh; and Patil, Manasi (2026) "Hybrid Model for Phishing Website Detection Using Transfer Learning," ASEAN Journal on Science and Technology for Development: Vol. 43: No. 1, Article 19.
DOI: https://doi.org/10.61931/2224-9028.1653
Available at: https://ajstd.ubd.edu.bn/journal/vol43/iss1/19