SS-CURL: Self-Supervised Contrastive URL Learning for Adversarially Robust Phishing Detection

Gattu, Srikanth (2025) SS-CURL: Self-Supervised Contrastive URL Learning for Adversarially Robust Phishing Detection. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (839kB) | Preview
Preview	PDF (Configuration Manual) Download (757kB) | Preview

Abstract

Phishing attacks are a big threat to cybersecurity, and the old ways of finding them don't work well against zero-day attacks and other ways that attackers try to hide. This research investigates the efficacy of self-supervised contrastive learning, combined with adversarial augmentation, in enhancing the reliable detection of advanced phishing URLs without relying on large labeled datasets. The SS-CURL (Self-Supervised Contrastive URL Learning) framework was developed by integrating transformer-based URL encoding with an innovative two-phase training approach. In the first phase, contrastive pre-training with URL-specific augmentations is used to learn how to tell the difference between unlabelled data. In the second phase, adversarial training with Projected Gradient Descent attacks is used during supervised fine-tuning. When tested on the PhiUSIIL dataset, which has 235,795 URLs, SS-CURL got an F1-score of 98.93% and an AUCROC score of 99.78%, which is as good as the best supervised methods. When the framework is attacked, its average accuracy only drops by 2.05%, which is a 15-fold improvement in robustness over traditional detectors, which lose 30% to 50% of their accuracy. The research brings forward a novel integration concept of contrastive learning with adversarial learning for tackling phishing tasks, a taxonomy for URL augmentation methods associated with authentic attack patterns, and insights confirming that pretraining methods utilizing self-supervised learning approaches can aid in improving adversarial tasks without affecting classification performance. The results are very helpful for making systems that can change and help deal with threats that are always changing.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Heffernan, Niall UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Computer software > Computer Security T Technology > T Technology (General) > Information Technology > Computer software > Computer Security Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Cyber Security
Depositing User:	Ciara O'Brien
Date Deposited:	03 Sep 2026 11:03
Last Modified:	03 Sep 2026 11:03
URI:	https://norma.ncirl.ie/id/eprint/9798

Actions (login required)

View Item