Enhancing Phishing Url Detection Resilience via Gan-Based Adversarial Training

Phishing remains one of the most pervasive cyber threats, with adversaries constantly adapting to bypass machine learning based URL detection systems. Despite impressive benchmark performance, often exceeding 99% accuracy, state-of-the-art classifiers are critically vulnerable to adversarial attacks. In this study, we propose an approach that enhances classifier robustness by incorporating adversarial training using synthetic phishing URLs generated by a Wasserstein Generative Adversarial Network. We train a baseline LSTM classifier and evaluate it under evasion attacks using handcrafted adversarial URLs. The proposed GAN is trained on real phishing URLs to generate synthetic samples that conform to URI syntax, enriching the training data and improving model resilience. Experimental results show that the adversarial training reduces attack success rates by 5% and improves classification accuracy under attack from 63.16% to 68.16%, with a corresponding increase in F1-score. This performance represents a significant improvement over prior studies and confirms that adversarially augmented training data enhances real-world effectiveness. The results confirm that incorporating synthetic phishing data through GAN-based adversarial training leads to measurable performance improvements, reducing vulnerability to evasion attacks and supporting more robust phishing detection in practice.