Visual-based Phishing Website Recognition
Tianjing Niu, Bin Ying Wu · 2024
Recently, there has been a growing trend towards vision-based interpretable phishing detection methods, which typically involve comparing suspicious webpages against a series of legitimate brand webpages. If the domain of a suspicious webpage does not align with those of the legitimate webpages, it is classified as a phishing website. In this paper, we propose a computer vision-based phishing identification framework. Initially, screenshots of suspicious webpages are input into a detection network, where a dual-path feature aggregation module is utilized to obtain multi-scale features with strong semantic and spatial prior, facilitating the detection of input boxes and logos. The logos are then input into a recognition network to generate feature vectors. To enhance the representational capacity of these features, dense feature aggregation and self-attention mechanisms are employed. Subsequently, the cosine similarity is computed to measure the similarity between suspicious logos and reference logos, and the domain names of suspicious logos are compared against those of reference logos to determine phishing websites. Experimental results demonstrate that our framework achieves state-of-the-art performance on the mixed dataset of Phishing Webpage dataset and Benign Webpage dataset, which shows 97.7% in precision, 95.0% in recall and 96.1% in F1-score. The code will be released soon.