Simulations of Event-Based Cyber Dynamics via Adversarial Machine Learning

Abstract

In this paper, we apply cybersecurity dynamics theory into practical scenarios. We use machine learning models as detection tools of intrusion detection systems and consider cyber attacks against node computers as well as adversarial attacks against machine learning models. We pay our attention to two problems. The first problem is when the network is attacked, how we can observe the states of the network and estimate its equilibrium with a lower cost. We apply an event-based observation and estimation method combined with machine learning-based intrusion detection systems. The second problem is to control the cost and the convergence speed of cybersecurity dynamics when it is under attack. An event-based control method and machine learning-based intrusion detection systems are put into use in this scenario. We simulate both scenarios and analyze the dynamics’ behaviors under an adversarial attack against the machine learning models on intrusion detection systems.