Research on Technology for Detecting Network Intrusions Based on Stacking Fusion Model

When it comes to identifying network intrusions, the label-based intrusion detection classification approach has been extensively utilized. Researchers have developed various network intrusion detection models employing a range of machine learning techniques; however, most existing models rely on a single classifier. This study develops an intrusion detection system that integrates multiple machine learning models. This study utilizes the NSL-KDD datasets as its primary focus of research. Firstly, an exploration of the datasets reveals issues related to class imbalance. To analyze the features within these datasets, several methods are employed, including variance analysis, chi-square test, Gini importance coefficient, and principal component analysis. Subsequently, this paper uses the grid search method and the cost-sensitive function method to train multiple logistic regression, decision tree, random forest and LightGBM models to determine the optimal hyperparameters for each model. Finally, this study establishes an intrusion detection model based on stacking fusion that utilizes LightGBM as the meta-learner while incorporating multiple logistic regression, decision tree, and random forest classifiers as base learners. Experimental results demonstrate that the efficacy of this fusion-based intrusion detection model surpasses that of individual classifiers. On both training and test sets, the constructed stacking fusion model achieves balanced accuracies of 98.94% and 97.21%, respectively.