Evaluation of ML Models using SMOTE and Feature Scaling for Intrusion Detection System(IDS)

This study compares the performance of K-Nearest Neighbors (KNN), Random Forest and XGBoost (XGB) classifiers across multiple feature sets for intrusion detection. MinMaxScaler and RobustScaler, two specific scalers, preprocessed the dataset and the Synthetic Minority Oversampling Technique (SMOTE) addressed the class imbalance. It also improved model performance by creating many more synthetic samples for the minority class and then calculated thorough performance metrics for both classes (0 and 1), including accuracy, precision, recall and F1 scores. Among the classifiers evaluated, the KNN classifier with MinMaxScaler achieved the best performance, exhibiting a 97.5% accuracy rate along with balanced precision, recall, as well as F1 scores of 0.98. RobustScaler was used and the accuracy of KNN was slightly lower 94.75% as a result. Random Forest and XGBoost models showed high accuracy along with balanced metrics, thus supporting SMOTE's effectiveness. SMOTE's application greatly improved the recall and Fl F1 scores for the minority class and this guaranteed substantially better detection of many underrepresented instances. This study shows the importance of SMOTE for dealing with imbalanced class sizes and improving classifier accuracy and indicates that MinMaxScaler is the optimal scaling method. The results considerably stress the large importance of combining SMOTE with three carefully selected preprocessing methods to achieve demonstrably optimal model performance in a meaningful number of imbalanced datasets.