Beyond Text-To-Image: Leveraging Natural Image Overlays for Robust Phishing Detection

Phishing attacks remain a significant threat to online security. While previous research has shown promising results in phishing detection using text-to-image conversion and deep learning models, limitations persist. This paper introduces a novel approach to enhance phishing detection by first converting text-based URL to images, then generating a composite image by overlaying natural images with URL images and finally training them on deep convolutional neural networks. We evaluate the performance of a pre-trained ResNet50 model, originally designed for natural image classification, on the composite dataset. Our experiments demonstrate the efficacy of our overlay approach, conducted through comparative analysis against baseline models trained solely on URL representations and against models incorporating URL-based blacklisting techniques. These comparisons highlight the unique contribution of the natural image overlay in enhancing detection accuracy beyond traditional text-based and blacklisting methods. This finding highlights the potential of leveraging pre-trained models and incorporating real-world visual cues to strengthen phishing detection systems.