METHODS AND MEANS TO IMPROVE THE EFFICIENCY OF NETWORK TRAFFIC SECURITY MONITORING BASED ON ARTIFICIAL INTELLIGENCE

Abstract

This paper aims to provide a solution for malicious network traffic detection and categorization. Remote attacks on computer systems are becoming more common and more dangerous nowadays. This is due to several factors, some of which are as follows: first of all, the usage of computer networks and network infrastructure overall is on the rise, with tools such as messengers, email, and so on. Second, alongside increased usage, the amount of sensitive information being transmitted over networks has also grown. Third, the usage of computer networks for complex systems, such as grid and cloud computing, as well as IoT and “smart” locations (e.g., “smart city”) has also seen an increase. Detecting malicious network traffic is the first step in defending against a remote attack. Historically, this was handled by a variety of algorithms, including machine learning algorithms such as clustering.

However, these algorithms require a large amount of sample data to be effective against a given attack. This means that defending against zero‑day attacks or attacks with high variance in input data proves difficult for such algorithms. In this paper, we propose a semi‑supervised generative adversarial network (GAN) to train a discriminator model to categorize malicious traffic as well as identify malicious and non‑malicious traffic. The proposed solution consists of a GAN generator that creates tabular data representing network traffic from a remote attack and a classifier deep neural network for said traffic. The main goal is to achieve accurate categorization of malicious traffic with a few labeled examples. This can also, in theory, improve classification accuracy compared to fully supervised models. It may also improve the model’s performance against completely new types of attacks. The resulting model shows a prediction accuracy of 91 %, which is lower than a conventional deep learning model; however, this accuracy is achieved with a small sample of data (under 1000 labeled examples). As such, the results of this research may be used to improve computer system security, for example, by using dynamic firewall rule adjustments based on the results of incoming traffic classification. The proposed model was implemented and tested in the Python programming language and the TensorFlow framework. The dataset used for testing is the NSL‑KDD dataset.