VAE-WACGAN: An Improved Data Augmentation Method Based on VAEGAN for Intrusion Detection.

Area of Science:

• Cybersecurity
• Machine Learning
• Data Science

Background:

• Class imbalance in network intrusion detection datasets degrades model performance.
• Existing methods struggle to generate realistic minority class samples for effective dataset balancing.

Purpose of the Study:

• To propose a novel generative model, VAE-WACGAN, for addressing class imbalance in network intrusion detection.
• To enhance the quality of generated minority class samples and improve training stability.

Main Methods:

• Developed VAE-WACGAN by integrating Variational Autoencoder Generative Adversarial Network (VAEGAN), Auxiliary Classifier Generative Adversarial Network (ACGAN), and Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP).
• Utilized VAE-WACGAN for oversampling anomalous data to generate realistic synthetic anomalies.
• Validated the approach on UNSW-NB15 and CIC-IDS2017 datasets.

Main Results:

• The VAE-WACGAN model generated high-quality synthetic anomalies closely mirroring actual network traffic distributions.
• Oversampling with VAE-WACGAN effectively balanced the network traffic datasets.
• Intrusion detection models trained on balanced datasets showed significantly enhanced performance metrics.

Conclusions:

• VAE-WACGAN effectively addresses class imbalance in network intrusion detection datasets.
• The proposed method improves the performance of intrusion detection models.
• VAE-WACGAN-based intrusion detection demonstrates superior effectiveness compared to other advanced methods in network security.