Strengthening Web Phishing Detection Against Adversarial Techniques

Phishing attacks pose significant threats by mimicking legitimate websites to steal user credentials. To address this, the study explores novel techniques for detecting phishing websites, focusing on evasion attacks that exploit vulnerabilities in existing classifiers. This research proposes a method combining URL feature extraction and machine learning classification to enhance phishing detection. The methodology includes crafting adversarial samples to test classifier robustness and evaluating phishing website appearance via image alteration metrics like mean squared error. The study utilizes the Phish Tank and Alex Data datasets, applying various classifiers with Logistic Deterioration, SVM, KNN, Decision Tree, Random Forest, and a Voting Classifier (Bagging with RF + DT). The Voting Classifier achieved accuracies of 96% and 85% on the Phish Tank and Alex Data datasets, respectively. Additionally, a Flask-based front end with user authentication will be developed for practical user testing. This research introduces a significant advancement in phishing detection and defense mechanisms.