Copula entropy regularization transformer with C2 variational autoencoder and fine-tuned hybrid DL model for network intrusion detection

Abstract

In cyber security, Intrusion Detection Systems (IDS) act as a network security tool, in which computational complexity and dynamic IDS detection issues are observed by conventional studies. In this paper, a novel Copula Entropy Regularization Transformer with$C^2$ variational autoencoder and Fine-tuned Hybrid Deep Learning (DL) model is introduced for Network Intrusion Detection. In previous IDSs studies, flow-based feature extraction techniques are concentrated, which is ineffective for detecting high-dimensional anomaly data. This work proposes a novel Copula Entropy Regularization Transformer with $C^2$ variational autoencoder for regularizing the feature extraction and feature selection, using a self-paced regularization mechanism. Recently, the pull towards IDS with Zero-Day (ZD) attacks gets increased, and the existing studies over it, possess high False-Negative Rates (FNR), leading to limited practical usage. For reducing the FNR and to improve the identification of ZD, a Fine-tuned Hybrid DL attack prediction model with deep Transudative Federated Transfer Learning (TFTL) is proposed. This gives out a map connection between known and zero-day attacks, data points for different dynamic network traffic, and classification of known and unknown network attacks. To investigate the ZD attack detection with the proposed model, it is validated in Python platform with Network Intrusion Detection dataset and the performance results show that the work gives better accuracy (98.54%), F1-score (97.5%), recall (97.302%), precision (98.2%) and detection rate of about 0.975%, while the comparative results show that this approach achieves a comparatively low false positive rate of 0.1, yielding high detection rate, and high accuracy in predicting attacks.