A Transformer network calibrated with fuzzy logic for phishing URL detection

Abstract

In the field of cybersecurity, phishing attacks by leveraging deceptive URLs continue to be a formidable challenge. These attacks evolve continuously, often rendering traditional detection methods inadequate. Even powerful deep learning models lack the adaptability required to keep pace with rapidly shifting phishing tactics. In this paper, we propose a novel fuzzy-calibrated transformer network for phishing URL detection. This model integrates a transformer network with the expert knowledge offered by fuzzy logic, enhancing its ability to interpret and adapt to the complex patterns of phishing URLs. This integration addresses the limitations of previous models, particularly their dependence on historical data, which is often outdated as phishing strategies evolve. Empirical evaluations of the proposed model on real-world datasets, which include over a million URLs, demonstrate its superior accuracy and adaptability in detecting phishing URLs, particularly in identifying novel and emerging phishing tactics. In experiments simulating real-world phishing detection scenarios, our model achieves an accuracy of 98.93 %, with a precision of 98.54 %, recall of 97.84 %, and F1-score of 98.03 %, outperforming baseline models by 5 %p in accuracy, highlighing adaptability to evolving phishing strategies.