Enhanced Intrusion Detection Using AI-Driven Data Balancing and VQ-VAE-Based Feature Extraction

Shivanthana S.^1,*, Manicka Raja M.², Lalitha Krishnasamy¹, Karthik R.¹, R. Venkatesan¹

¹Division of Computer Science and Engineering, Karunya Institute of Technology and Sciences, Coimbatore

²Department of Artificial Intelligence and Data Science, Nandha Engineering College, Erode

Emails: shivanthanas@karunya.edu.in; manickaraja@karunya.edu; lalithak@nandhaengg.org; karthikr@karunya.edu; rlvenkei2000@gmail.com

Abstract

Network security faces significant challenges due to the increasing sophistication of cyber threats and the inherent class imbalance in intrusion detection datasets. To address this issue, a hybrid Boundary Equilibrium Generative Adversarial Network (BEGFAN) and Vector Quantization Variational Autoencoder (VQVAE) framework, termed BVQVAE, is proposed for Network Intrusion Detection Systems (NIDS). The framework involves preprocessing, feature extraction, and class balancing to enhance classification accuracy. Missing values are imputed, categorical features are label-encoded, and numerical attributes are normalized to ensure a structured dataset. BEGAN generates synthetic samples to mitigate class imbalance, while VQVAE extracts essential features using an encoder with quantization and a decoder for network traffic reconstruction. The model is evaluated on NSL-KDD and UNSW-NB15 datasets, achieving 82.56% accuracy, with precision, recall, G-mean, and F1-score of 86.53%, 87.65%, 86.21%, and 87.08%, respectively.

Keywords: Network Security; Class Imbalance; Adversarial Learning; Anomaly; Variational Autoencoder