Adapted CNN-SMOTE-BGMM Deep Learning Framework for Network Intrusion Detection using Unbalanced Dataset

Waad F.  Kamil; mad J.  Mohammed

doi:10.24996/ijs.2023.64.9.43

Authors

Waad F. Kamil Department of Computer Science, College of Science, University of Baghdad, Baghdad, Iraq
mad J. Mohammed Department of Computer Science, College of Science, University of Baghdad, Baghdad, Iraq https://orcid.org/0000-0002-5829-6153

DOI:

https://doi.org/10.24996/ijs.2023.64.9.43

Keywords:

Bayesian Gaussian Mixture Model, Convolutional Neural Network, Deep Learning, Extreme Gradient Boosting, Machine Learning, Recursive Feature Elimination, Synthetic Minority Oversampling Technique

Abstract

This paper proposes a method for improving network security by introducing an Intrusion Detection System (IDS) framework between end devices. It considers big data challenges and the difficulty of updating databases to uncover new threats to firewalls and detection systems. The proposed framework introduces a supervised network using CNN and the UNSW-NB15 dataset. Recursive Feature Elimination (RFE) and Extreme Gradient Boosting ( XGB)wereused to save time and resources. The Synthetic Minority Oversampling Technique (SMOTE) and Bayesian Gaussian Mixture Model (BGMM) reduce bias toward the majority class of the dataset. The results show that this model performs better than other methods, with 98.80% accuracy for binary classification and 96.49% for classification into multiple groups.