Detecção de intrusão em sistemas ciber-físicos com uso de técnicas de aprendizado de máquina

Abstract

Cyber-Physical Systems (CPS) are increasingly interconnected to the most diverse critical infrastructure systems that can impact human life in our daily lives. The security of digital information is of vital importance for these systems, such as industrial water treatment plants or Smart Grid electrical grids, where an attack can harm a part of the distribution of vital resources or electrical equipment if the necessary security measures are not adopted. In this context, Intrusion Detection Systems (IDS) are needed as a second line of defense for these systems that are so important in our daily lives. The use of Machine Learning (ML) techniques in IDS is important as an automatic way to detect intrusions with minimal human intervention. In this context, machine learning techniques on standard datasets are known as Offline ML. However, with the increasing need to use IDS in Big Data with time constraints and limited memory, there is a need to use data flow machine learning techniques, which are known as Online ML. In this last technique, each sample must pass the learning and testing stages only once, without being stored in memory, due to the large flow of data. Many studies in the area of IDS in CPS do not make adequate use of specific datasets of the area representing the actual existing equipment of these systems. This motivation led to a survey of the main datasets existing in the literature referring to CPS that will be later used in the experiments of this dissertation. Therefore, this work aims to compare the intrusion detection in cyber-physical systems with these suitable datasets using IDS methods with Offline Machine Learning techniques and Online Machine Learning.