Prüfungssenat:
Univ.-Prof.Dr. Tanja Zseby / E 389 (BetreuerIn)
Prof. Paul Smith / Lancaster University
Prof. Norbert Pohlmann / Westfälische Hochschule
ao.Univ.-Prof.Dr. Thilo Sauter / E 384 (Vorsitz)
Prüfungssenat:
Univ.-Prof.Dr. Tanja Zseby / E 389 (BetreuerIn)
Prof. Paul Smith / Lancaster University
Prof. Norbert Pohlmann / Westfälische Hochschule
ao.Univ.-Prof.Dr. Thilo Sauter / E 384 (Vorsitz)
Abstract
Intrusion detection systems (IDS) offer significant advantages for industrial control system (ICS) networks by providing real-time insight into data flows, enabling the detection of malicious network behavior, identifying previously unknown hosts, and allowing ongoing attacks to be detected and responded to. Anomaly-based approaches are particularly suited to ICS networks, as network traffic in such environments is typically far more deterministic, and infrastructure changes occur far less frequent than in conventional IT networks. Anomaly-based approaches are particularly suited to ICS networks, as network traffic in such environments is typically far more deterministic, and infrastructure changes occur far less frequent than in conventional IT networks.Anomaly-based approaches are particularly suited to ICS networks, as network traffic in such environments is typically far more deterministic, and infrastructure changes occur far less frequent than in conventional IT networks.However, many modern commercial IDS rely primarily on signature-based detection and therefore fail to exploit useful, domain specific characteristics of ICS traffic. In this dissertation, we investigate the potential of machine-learning-based IDSs in ICS networks. We develop, evaluate, and demonstrate an approach for reliably detecting novel or concealed attacks in industrial control systems (ICS) networks. We first identify suitable machine learning classifiers and parameters for reliable anomaly detection based on the current state of scientific literature. Since encryption has become the norm in ICS networks, we limit our methods to those that can handle encrypted traffic. We include findings from previous work at our institute on feature selection for encrypted traffic and use practical experiments to compare methods and determine optimal parameter settings. Based on these findings, we select a set of candidates for further investigation: Random forest, one-class support vector machine, two different autoencoders, and clustering. Next, we develop a lab prototype to compare the chosen classifier candidates under controlled conditions. For the evaluation, we use a public IT dataset and our own recorded ICS datasets, obtained at the TU Wien pilot factory in Vienna, Austria. We then investigate the relevant relations between security and safety. Safety is a primary concern in ICS networks, and security and safety measures may conflict. We then create a method to assess potential safety consequences of attacks in an automated way. In the next step, we develop the second prototype, the safety-aware prototype, which integrated safety considerations into the IDS. Our method can evaluate the feasibility of accounting for safety consequences. Our method assigns a priority score to each security-related incident, enabling operators to prioritize and act more efficiently when multiple incidents occur simultaneously. Lastly, we design a practically deployable live prototype: A neural network learns the typical network traffic for the specific network over a defined period of time. During training, this IDS tailors itself to the specific network it is used in and can alert if a host shows a significant behavior change. This way, it can reliably detect both known and unknown attacks, including those in encrypted network traffic.