Presenting A Hybrid Method of Deep Neural Networks to Prevent Intrusion in Computer Networks
Subject Areas : Intelligent Multimedia Processing and Communication Systems (IMPCS)
Mohsen
Roknaldini
1
(Department of Computer Engineering, Qeshm branch, Islamic Azad university, Qeshm, Iran,)
Erfaneh
Noroozi
2
(Department of Computer Engineering, Qeshm branch, Islamic Azad university, Qeshm, Iran,)
Keywords: Deep learning, neural network, feature selection, intrusion detection system,
Abstract :
Introduction: Nowadays, computer networks have significant impacts on our daily lives, leading to cybersecurity becoming a crucial area of research. Cybersecurity techniques mainly encompass antivirus software, firewalls, and intrusion detection systems. Intrusion detection system is one of the fundamental security tools in the field of computer networks and systems. The primary goal of an intrusion detection system is to identify and alert about any unauthorized activities, threats, or attacks on a system or network. By analyzing the flow of data and network/system events, the intrusion detection system attempts to identify patterns and indicators related to various attacks and intrusions. Intrusion detection systems can operate based on rules or learning. In the rule-based approach, algorithms and rules created by security experts and analysts are used to detect patterns and identify attacks. However, in the machine learning approach, machine learning algorithms and deep neural networks are employed to extract patterns and features related to attacks from real data. Method: This study focuses on the examination and presentation of a combined approach using deep neural networks to prevent intrusions in computer networks. The primary objective of this research is to enhance the efficiency of intrusion detection systems. To achieve this goal, a combined approach of deep learning and artificial neural networks is proposed. This approach utilizes deep neural networks to detect more complex features and improves the model's performance. Results: Simulation results demonstrate that deep neural network methods such as MLP, CNN, LSTM, and GRU yield favorable outcomes compared to other single-layer machine learning techniques. In this study, two combined methods, CNN-GRU and CNN-LSTM, were introduced and tested on the KDD CUP'99 dataset for comprehensive analysis and evaluation. Both combined approaches exhibit high accuracy and lower classification errors compared to other introduced methods. Therefore, it can be concluded that the CNN-LSTM combined approach performs well on the KDD CUP'99 dataset. Discussion: Based on the achieved results, the combined CNN-LSTM and CNN-GRU methods offer very good performance with accuracies of 99.95% and 99.92%, respectively, on the KDD CUP'99 dataset. Among these methods, minor differences in the performance of some parameters for classes may exist, yet both approaches remain acceptable. Hence, it can be concluded that the combined CNN-LSTM approach performs well on the KDD CUP'99 dataset.
