A New Approach for Detecting Network Intrusion Based on Anomalies Using a Deep Clustering Variational Auto-Encoder

Semi-supervised network intrusion detection systems are becoming more vital in today's fast-evolving digital ecosystem. While increasing interest in commercial and academic contexts is rising, specific accuracy difficulties still need to be resolved. Two significant challenges contributing to this fear are accurately learning the probability distribution of standard network data and identifying the boundary between normal and abnormal data locations in the latent space. Several methods have been proposed for semisupervised learning of the latent representation of standard data, including clustering-based Autoencoders (CAEs) and hybridized approaches combining Principal Component Analysis (PCA) and CAEs. Inadequate handling of highdimensional data and excessive dependence on feature engineering remain limitations of current methods. To combat these problems and boost the efficiency of network intrusion detection, we introduce a novel deep learning model called Cluster Variational Autoencoder (CVAE). This approach
allows for a more condensed and dominant representation of
the latent space. Thanks mainly to the VAE's ability to
comprehend the fundamental probability distribution of
specific network data, we have broken through these barriers.
The proposed model is tested on eight different network
intrusion benchmark datasets. These datasets include NSLKDD, UNSW-NB15, and CICIDS2017. Experimental findings demonstrate that our method outperforms state-ofthe-art semi-supervised methods.