Applying machine learning techniques on the detection of cyber-physic attacks
Costeira, Ivo Micael Frazão
MSc Thesis · 2018 · University of Coimbra
Supervisors: Helder de Jesus Araújo, Tiago José dos Santos Martins da Cruz
Abstract
Today’s society has had an increasing reliance on Industrial (Automation and) Control Systems, especially in the control of Critical Infrastructures, such as generation, transmission and distribution of electrical energy, or water treatment, and where the uninterrupted work of the infrastructure is essential for the safety and livelihood of a modern society. A disruption of the normal operation of such infrastructures, due to a fault or an attack, has the potential to create serious consequences, whereby these systems have strategic significance, which cannot be ignored. These infrastructures were designed with the assumption that an airgap from the outside world was in place, however, their increasing reliance on Commercial, Off-The-Shelf equipment, open technologies and open protocols has exposed them to threats that previously were not a concern.Several studies have proposed many strategies to secure these infrastructures, which mainly rely on novel algorithms or domain-adaptations of known algorithms. However, the constraints of these infrastructures prevent the existence of common datasets from which a comparison of their performances can be made. This work aims at presenting some insight on how many of the most used algorithms in the literature behave when some of the characteristics of datasets are varied, namely the size of the traffic captures and the relative size of the attacks within those captures. An insight on the effects of hyperparameter tuning of the algorithms is also studied. Finally, a generalized framework for the detection of Man-in-the-Middle attacks is presented based on the Time-To-Response of the TCP packets within the network.