Open Access

ARTICLE

Deep Learning Network Intrusion Detection Based on MI-XGBoost Feature Selection

Manzheng Yuan^1,2, Kai Yang^2,*

1 School of Computer, Xijing University, Xi’an, 710123, China
2 Xi’an Key Laboratory of Human-Machine Integration and Control Technology for Intelligent Rehabilitation, School of Computer Science, Xijing University, Xi’an, 710123, China

* Corresponding Author: Kai Yang. Email:

Journal of Cyber Security 2025, 7, 197-219. https://doi.org/10.32604/jcs.2025.066089

Received 29 March 2025; Accepted 11 June 2025; Issue published 07 July 2025

Abstract

Currently, network intrusion detection systems (NIDS) face significant challenges in feature redundancy and high computational complexity, which hinder the improvement of detection performance and significantly reduce operational efficiency. To address these issues, this paper proposes an innovative weighted feature selection method combining mutual information and Extreme Gradient Boosting (XGBoost). This method aims to leverage their strengths to identify crucial feature subsets for intrusion detection accurately. Specifically, it first calculates the mutual information scores between features and target variables to evaluate individual discriminatory capabilities of features and uses XGBoost to obtain feature importance scores reflecting their comprehensive contributions in complex data relationships. Then, through adaptive weighted combination of the two types of scores to generate integrated feature weights, we can select the most valuable features for intrusion detection based on these weights. Furthermore, a deep learning detection model combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM) is developed based on the selected features. This model takes advantage of CNN’s capability in extracting local feature associations and BiLSTM’s strength in capturing long-term dependencies in sequential data, further enhancing the accuracy and efficiency of intrusion detection. Experiments have shown that on the NSL-KDD and UNSW-NB15 datasets, the accuracies reach 99.35% and 88.78%, respectively, demonstrating that this method significantly improves the accuracy and efficiency of network intrusion detection. This research provides new ideas for feature selection and model construction of network intrusion detection systems. It can improve the detection performance while reducing the computational overhead, and helps optimize the practical application effect of the IDS, having certain technical reference value for enhancing the intelligent level of network intrusion detection.

---

Keywords

---