Adaptive Phishing URL Classification: A Generative Adversarial Approach

Phishing detectors trained on old data quietly lose their edge as attackers change tactics. This paper trains a generator to invent realistic new phishing URLs and pits it against a detector, so the detector learns to handle attacks it has never seen. Tested against a private feed of current phishing links from 2024–2025, the adversarially trained detector caught about 5% more modern phishing URLs, showing this cat-and-mouse training helps classifiers keep up with evolving threats.