Deep Learning Algorithms for Real-time Antivirus Systems: A New Approach to Detect and Mitigate Cyber Threats

Abstract

The purpose of the study was to develop methods for integrating deep learning algorithms into antivirus systems to increase the effectiveness of detecting and neutralising cyber threats. Modern deep learning algorithms were analysed and tested, including convolutional neural networks, variational autoencoders, generative-adversarial networks, recurrent neural networks, and transformers. Each of the algorithms has been adapted for the tasks of detecting and neutralising threats, and integrated into prototypes of antivirus systems. The main results included antivirus platforms such as CrowdStrike, Sophos, Darktrace, SentinelOne, and CylancePROTECT. Convolutional neural networks have shown high efficiency in analysing programme behaviour and classifying network traffic, which makes them suitable for identifying known and new threats. Autoencoders have demonstrated their effectiveness in detecting anomalies, which increases the accuracy of recognising rare and unknown attacks. Generative-adversarial networks have made it possible to model threat scenarios and improve the adaptability of antivirus systems. Recurrent neural networks with their ability to analyse time dependencies have provided accurate predictions based on programme behaviour. Transformers proved to be the most effective for analysing the structure of the code, which made it possible to identify complex malware. In other words, the results of the study confirm that the implementation of deep learning algorithms in antivirus systems such as Darktrace and SentinelOne provides a higher level of protection and adaptability. Sophos and CylancePROTECT, due to the use of autoencoders and transformers, have shown their effectiveness in detecting complex threats with a minimum number of false positives. Thus, the proposed approach expands the capabilities of modern antivirus systems, making them more reliable and flexible in the face of rapidly changing cyber threats.