Comparison of SVM and Gradient Boosting with PCA for Website Phising Detection

Abstract

The increasing use of the internet has led to a rise in phishing attacks, posing a threat to user data security. This study compares the performance of the Support Vector Machine (SVM) and Gradient Boosting algorithms, integrated with Principal Component Analysis (PCA) for dimensionality reduction, in classifying phishing websites. The dataset consists of 11,054 samples classified into two categories: phishing (1) and non-phishing (-1), with three data partition scenarios for training and testing: 70:30, 80:20, and 90:10. Experimental results indicate that SVM outperforms Gradient Boosting in terms of accuracy and recall, particularly in detecting phishing websites. In the 80:20 and 70:30 data partition scenarios, the SVM model achieved an accuracy of 96% to 97% and had a higher recall for phishing websites, making it more sensitive to phishing detection. However, Gradient Boosting demonstrated consistent performance with an accuracy of around 94%, providing a balanced result between precision and recall for both classes. Therefore, the SVM model is superior for phishing detection tasks requiring high sensitivity to phishing websites, while Gradient Boosting remains a viable alternative when a more balanced performance between phishing and non-phishing sites is needed. The study concludes that both algorithms can be effectively used for phishing detection, with potential improvements through further experiments and hyperparameter tuning.