A high-accuracy phishing website detection method based on machine learning

Abstract

The rapid development of e-commerce, e-banking, and social networks has made phishing attack detection one of the most critical technologies in all cyber security systems. To improve the efficiency of anti-phishing techniques, we present an improved predictive model based on machine learning. The proposed method uses six different algorithms; Logistic Regression, K-Nearest Neighbors, Naive Bayes, Random Forest, Support Vector Machine, and Extreme Gradient Boosting (XGBoost). The experiments are based on a public dataset of 58,000 legitimate websites and 30,647 phishing ones, including 112 attributes for each sample. Our evaluations in the feature selection process show that after balancing the dataset and dropping constant features, a noticeable improvement can be achieved. We conducted our evaluation found on eight major unique scenarios. The experimental results of our phishing websites detection (PWD) method indicate remarkable performances in which each algorithm reached an accuracy of more than 93%, and the XGBoost classifier outperforms others with 99.2% overall accuracy, 99.1% precision, 99.4% recall, and 99.1% specificity. In addition, the study achieved optimal run-time of about 1500 ms for the XGBoost algorithm without dimension reduction while using Principal Component Analysis (PCA) reduces it down to just 869 ms. As a result, the proposed approach would be practical in both offline and real-time applications.