Model of automated attack simulation system based on machine learning

M. M. Grekov

Tula State University, 92 Lenina Ave., Tula 300012, Russian Federation

Abstract: The paper explores the application of machine learning techniques to the development of an intelligent attack simulation system designed to automate penetration testing as a part of network security assurance. Traditional penetration testing methods are often labor-intensive and require significant time and resource investment. The proposed model integrates Generative Adversarial Imitation Learning and reinforcement learning algorithms, enabling the simulation of attacker behavior and the generation of realistic scenarios that closely mimic the actions of professional security experts. A key feature of the model is the incorporation of semantic rewards which account not only for the achievement of attack objectives but also for factors such as the novelty and stealthiness of the actions. To improve adaptability in dynamic network environments, the model can be extended with dual discriminators. Additionally, support for multi-agent interaction makes it possible to simulate coordinated attacks involving multiple adversaries.

Keywords: information security, generative adversarial imitation learning, attack simulation, machine learning.

Received: 30.01.2025
Accepted: 15.09.2025

DOI: 10.14357/08696527250307