A novel bagging- XGBoost ensemble model for attaining high accuracy and computational efficiency in network intrusion detection

Abstract

The study focuses on enhancing network intrusion detection to enhance network security and prevent potential data breaches. We propose B-XGBoost, an ensemble learning model that combines bagging and boosting, using 10k cross-validation and Bayesian optimization for binary network intrusion classification. The proposed model was trained and tested on the CIC-ID2017 dataset. Decision Trees, Random Forests, Support Vector Machines, Naive Bayes, k-Nearest Neighbors, and Neural Networks were trained and tested on the same dataset for performance comparison purposes. The results show that the B-XGBoost algorithm had the highest F1 Score (0.982), Precision (0.975), Recall (0.990), Cohen's Kappa (0.978), and ROC AUC (0.983). The other algorithms had varying levels of performance, with the Decision Trees having the second-highest F1 Score (0.950). Bayesian optimization significantly reduced the time, computational efficiency, and cost of hyperparameter tuning by using a probabilistic model to predict hyperparameters that resulted in high performance. The high scores in F1, precision, recall, agreement with human annotators, and ability to distinguish between positive and negative instances demonstrate the effectiveness of this approach in enhancing network security. For the best results of the B-XGBoost to be obtained, the hyperparameters of the base model need to be tuned to achieve maximum computational efficiency in light of the available resources.