Detection of Phishing Websites using Feature Selection and Image Embeddings

Phishing, a cyber threat originating in 1996, has evolved into a sophisticated and pervasive attack vector that exploits deceptive emails and counterfeit websites to steal sensitive information. Despite various preventive and detection strategies, a comprehensive solution remains limited. This study explores the application of machine learning to enhance phishing detection using a dataset of both phishing and legitimate URLs. We evaluate multiple models, including Decision Tree, Linear Regression, Random Forest, Naive Bayes, Gradient Boosting, KNN, and Support Vector Classifier. Additionally, we introduce a visual comparison technique using image embeddings and cosine similarity to refine detection accuracy, particularly for borderline cases. Performance is measured using accuracy, precision, recall, F1-score, and cosine similarity.