Network Intrusion Detection System Using Stacked Ensemble Model with SVM Smote Oversampling and Recursive Feature Elimination

Abstract

Intrusion Detection Systems (IDS) are essential for safeguarding networks against malicious activities, but traditional IDS models struggle with challenges such as class imbalance, high false alarm rates, and poor generalization. While machine learning (ML)-based IDS offer improvements, single classifier models suffer from bias, variance, and limited robustness. To address these limitations, this study proposes a Non-evolutionary Feature Selection-based Network Intrusion Detection System using Stacked Ensemble Learning (NFSNIDS). The proposed workflow begins with data preprocessing, where SVM SMOTE oversampling balances class distribution, Local Outlier Factor (LOF) outlier detection removes anomalies, Recursive Feature Elimination (RFE) selects relevant features, and Robust Scaler ensures effective data normalization. The processed data is then fed into a Stacked Ensemble Learning model comprising Extreme Gradient Boosting (XGB) and Extra Trees (ET) as base classifiers. Their outputs are used to create a new training set for a meta-classifier, which is trained using Logistic Regression to enhance predictive performance. The model is validated using 10-fold cross-validation, with Accuracy and F1-score as key performance metrics. Comparative evaluations against single classifiers, existing ensemble models, and benchmark IDS solutions confirm that NFSNIDS consistently outperforms all alternatives, making it a highly effective and robust approach for network intrusion detection.