Tuning of Anomaly Detectors in the Presence of Sensor Attacks
Time: Fri 2019-10-11 10.00
Location: V2, Teknikringen 76, Stockholm (English)
Subject area: Electrical Engineering
Doctoral student: David Umsonst , Reglerteknik
Opponent: Assistant Professor Justin Ruths, University of Texas at Dallas, USA
Supervisor: Professor Henrik Sandberg, Reglerteknik
Critical infrastructures, such as the power grid and water distribution networks, are the backbone of our modern society. With the integration of computational devices and communication networks in critical infrastructures, they have become more efficient, but also more vulnerable to cyberattacks. Due to the underlying physical process, these cyberattacks can not only have a financial and ecological impact, but also cost human lives. Several reported cyberattacks on critical infrastructures show that it is vital to protect them from these attacks. Critical infrastructures typically rely on accurate sensor measurements for optimal performance. In this thesis, we, therefore, look into attacks that corrupt the measurements.
The first part of the thesis is concerned with the feasibility of a worst-case sensor attack. The attacker's goal is to maximize its impact, while remaining undetected by an anomaly detector. The investigated worst-case attack strategy needs the exact controller state for its execution. Therefore, we start by looking into the feasibility of estimating the controller state by an attacker that has full model knowledge and access to all sensors. We show that an unstable controller prevents the attacker from estimating the controller state exactly and, therefore, makes the attack non-executable. Since unstable controllers come with their own issues, we propose a defense mechanism based on injecting uncertainty into the controller. Next, we examine the confidentiality of the anomaly detector. With access to the anomaly detector state, the attacker can design a more powerful attack. We show that, in the case of a detector with linear dynamics, the attacker is able to obtain an accurate estimate of the detector’s state.
The second part of the thesis is concerned with the performance of anomaly detectors under the investigated attack in the first part. We use a previously proposed metric to compare the performance of a χ2, cumulative sum (CUSUM), and multivariate exponentially weighted moving average (MEWMA) detectors. This metric depends on the attack impact and average time between false alarms. For two different processes, we observe that the CUSUM and MEWMA detectors, which both have internal dynamics, can mitigate the attack impact more than the static χ2 detector. Since this metric depends on the attack impact, which is usually hard to determine, we then propose a new metric. The new metric depends on the number of sensors, and the size of an invariant set guaranteeing that the attack remains undetected. The new metric leads to similar results as the previously proposed metric, but is less dependent on the attack modeling. Finally, we formulate a Stackelberg game to tune the anomaly detector thresholds in a cost-optimal manner, where the cost depends on the number of false alarms and the impact an attack would cause.