Zoneout regularization-gated recurrent unit algorithm on NIDS with class imbalance handling

Abstract

Network intrusion detection system (NIDS) is primarily utilized tool to identify malicious threats on the network. It plays an essential role in safeguarding against an increasing variety of attacks and ensures enhanced security for the network. The existing model struggled to handle the imbalance of class issues during the process of classification due to their biased nature, which reduced the performance of the algorithm. In this paper, the zoneout regularization–gated recurrent unit (ZR-GRU) algorithm is developed to detect and classify intrusions in the network. Incorporating the ZR into GRU reduces overfitting by preventing the model from becoming overly dependent on specific features. It provides good generalization by maintaining diversity in learned representation. Synthetic minority oversampling technique (SMOTE) and Near Miss methods are utilized to balance the samples in the dataset, which helps to increase the performance of a classifier in NIDS. The ZR-GRU technique attained 99.91% accuracy on UNSW-NB15, 99.92% accuracy on CIC-IDS2018, and 99.14% accuracy on CIC-DDoS2019 when comparing with a convolutional neural network bidirectional long short-term memory (CNN-BiLSTM).