Industrial Internet Intrusion Detection Method Based on VAE-WGAN-GP Data Enhancement
<p>Yahui Wang, Zhiyong Zhang</p> · Academic Journal of Computing & Information Science · 2024
Deep learning has played a significant role in intrusion detection. However, deep learning-based intrusion detection methods require a large amount of annotated data for model training. In the real world, the types of intrusion data that are of concern often belong to minority classes that lack labels. This imbalance creates an imbalanced dataset for intrusion detection, where normal data significantly outweighs attack data. Class imbalance can lead to biased decision boundaries, resulting in increased classification errors for attack data. In the face of imbalanced data, we propose a data augmentation model based on VAE-WGAN-GP. VAE-WGAN-GP combines variational autoencoders (VAE) and Wasserstein generative adversarial networks (WGAN) with gradient penalty (GP), creating a deep learning generative model. We augment the minority class data using this model to balance the dataset. Finally, we demonstrate significant improvements in multi-class intrusion detection using multiple classifiers by applying our data augmentation model to a traditional internet dataset and an industrial control system network dataset.