A Novel Phishing Website Detection Model Based on KMeansSMOTERENN and Stacking

Real-word datasets of anomaly detection often have neither symmetric nor normal but rather skewed and imbalanced class distribution, with the Majority class examples overwhelming the Minority class (es) examples [20]. However, the traditional classification algorithms consider datasets to have equal representation of each class and, therefore, equal misclassification costs [1-2], [4]. [20-22]. Both of these assumptions fall flat with class imbalanced [1-2], [4], [20-22]. Researchers have proposed several techniques to handle the class imbalance at the Data and Algorithm levels [1-2], [4], [20-22]. We propose a novel Phishing Website Detection Model based on ‘KMeansSMOTERENN’ hybrid sampling and Stacking, which is a fusion of three techniques, namely, (i) ‘KMeansSMOTERENN’ for Dataset balancing, (ii) Cost-sensitive Classification for Algorithm-balancing, and (iii) Hybrid Classification using Stacked Generalization. At Level 0 of our Hybrid Classification Model based on Stacking, we have employed Cost-Sensitive Decision Tree (DT), Cost-Sensitive Support Vector Machine (SVM), and Gaussian Naïve Bayes (GNB). At Level 1 of our Hybrid Classification Model based on Stacking, we have employed a Cost-Sensitive Logistic Regression (LR) Classifier. Through our extensive experiments with 10-fold Stratified Cross Validation X 10 runs, we have found that our novel proposed ‘KMeansSMOTERENN-STK’ Stacking model with Mean(AUROC) as 99.81 % and Mean(Accuracy) as 98.96% outperformed the ‘Original-STK’ Stacking model based on Original(Imbalanced) dataset with Mean(AUROC) as 99.24 % and Mean(Accuracy) as 96.57% We also understand that our novel proposed Phishing Website Stacking model based on KMeansSMOTERENN Hybrid Sampling and Cost-Sensitive Classification has not been proposed so far.