Adversarial threat evaluation of machine learning-based phishing detectors: a cybersecurity-focused black-box perspective
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Phishing attacks often succeed through subtle URL modifications that evade detection. We evaluated five URL-based classifiers under a label-preserving black-box attacker with a fixed edit budget across three scenarios: unaltered data, edited phishing URLs, and mixed streams. Model reasoning was examined using permutation-based importance, TreeSHAP, and Kernel-LIME. On clean data, ensemble models achieved F1 scores of approximately 0.981, MLP reached 0.978, and linear models attained 0.971. Under maximum perturbation (k = 9), XGBoost and MLP maintained F1 above 0.975, while other models declined slightly. Precision remained stable; decreases reflected missed detections. A lightweight augmentation technique improved adversarial recall by 0.003 to 0.007 without reducing precision. Attribution analysis showed ensemble models shifted toward fragile features as edits increased, and growing SHAP-LIME divergence may indicate stability drift during deployment.
Fields of Study & Subjects
Spam and Phishing Detection
Adversarial Robustness in Machine Learning
Hate Speech and Cyberbullying Detection
Publication details
- Journal / source
- International Journal of Information and Computer Security
- Volume
- 29
- Issue
- 4
- Pages
- 436–450
- Publication date
- 1 January 2026
- Language
- EN