Design of Network Intrusion Detection System Using Lion Optimization-Based Feature Selection with Deep Learning Model

Abstract

In the domain of network security, intrusion detection systems (IDSs) play a vital role in data security. While the utilization of the internet amongst consumers is increasing on a daily basis, the significance of security and privacy preservation of system alerts, due to malicious actions, is also increasing. IDS is a widely executed system that protects computer networks from attacks. For the identification of unknown attacks and anomalies, several Machine Learning (ML) approaches such as Neural Networks (NNs) are explored. However, in real-world applications, the classification performances of these approaches are fluctuant with distinct databases. The major reason for this drawback is the presence of some ineffective or redundant features. So, the current study proposes the Network Intrusion Detection System using a Lion Optimization Feature Selection with a Deep Learning (NIDS-LOFSDL) approach to remedy the aforementioned issue. The NIDS-LOFSDL technique follows the concept of FS with a hyperparameter-tuned DL model for the recognition of intrusions. For the purpose of FS, the NIDS-LOFSDL method uses the LOFS technique, which helps in improving the classification results. Furthermore, the attention-based bi-directional long short-term memory (ABiLSTM) system is applied for intrusion detection. In order to enhance the intrusion detection performance of the ABiLSTM algorithm, the gorilla troops optimizer (GTO) is deployed so as to perform hyperparameter tuning. Since trial-and-error manual hyperparameter tuning is a tedious process, the GTO-based hyperparameter tuning process is performed, which demonstrates the novelty of the work. In order to validate the enhanced solution of the NIDS-LOFSDL system in terms of intrusion detection, a comprehensive range of experiments was performed. The simulation values confirm the promising results of the NIDS-LOFSDL system compared to existing DL methodologies, with a maximum accuracy of 96.88% and 96.92% on UNSW-NB15 and AWID datasets, respectively.