Controller Cyber-Attack Detection and Isolation

Anna Sztyber-Betley, Michal Syfert, Jan Maciej Kóscielny, Zuzanna Górecka

TL;DR

This paper introduces a novel approach to detect and isolate cyber-attacks targeting industrial control systems (ICS) by integrating controller modeling with control loop performance indicators. This method offers a more holistic security layer by analyzing standard operational data, making it applicable to a wide range of industrial processes without requiring specific process models.

What Problem Does This Paper Address?

Industrial control systems (ICS) are increasingly vulnerable to cyber-attacks, which can have devastating consequences for critical infrastructure. While traditional cybersecurity measures protect network perimeters, sophisticated attacks can still penetrate and manipulate control algorithms, leading to system failures, safety hazards, and operational disruptions. Existing methods for fault detection and isolation (FDI) primarily focus on physical components and sensors, often overlooking the integrity of the control algorithms themselves. This paper addresses the critical need for a unified approach that can detect and isolate cyber-attacks specifically targeting the controller's operation, complementing existing FDI techniques.

Methods

The authors propose an integrated approach that combines two main strategies:

  • Controller Modeling: They employ both linear and neural network models to estimate the controller's behavior. By comparing the actual controller output with the model's prediction, they can detect anomalies that might indicate an attack. This approach is particularly useful for identifying changes in controller parameters or operation modes.
  • Control Loop Performance Indicators: A set of carefully selected indicators (e.g., mean absolute control error, control signal variance, saturation index) are used to assess the overall health and performance of the control loops. Changes in these indicators can signal deviations from normal operation, potentially caused by cyber-attacks.

These two approaches are combined through a preprocessing pipeline that includes calculating indicators in moving windows, applying Box-Cox transformations for non-normal data, and standardizing signals. The resulting binarized diagnostic signals are then used in a binary diagnostic matrix to isolate the specific type of cyber-attack. This matrix maps the observed diagnostic signals to known attack signatures, allowing for identification and localization of the threat.

Datasets

The authors generated a dataset for their case study using a simulated superheater system. This dataset includes normal operation data and data from various cyber-attack scenarios. The dataset is made publicly available at https://doi.org/10.5281/zenodo.7612269.

Key Insights and Ideas

  • Holistic Security Layer: By integrating controller modeling with performance indicators, the approach provides a more comprehensive security solution than traditional FDI methods, which often neglect the controller itself.
  • Data-Driven and Model-Agnostic: The method relies on standard operational data (PV, SP, CV) and does not require detailed, specific models of the controlled process, making it broadly applicable.
  • Distinguishing Cyber-Attacks from Faults: The approach aims to differentiate between physical faults and cyber-attacks by analyzing how they manifest in controller behavior and loop performance.
  • Importance of Controller Integrity: The research highlights that the controller's internal operation is a critical point of vulnerability and a key area for monitoring.

Key Results

  • High Detection Rates: The proposed method achieved a high detection percentage (99.09%) for attacks directly targeting controllers. For other types of attacks, the detection rate was 84.99%.
  • Accurate Isolation: The system demonstrated accurate isolation of control loops affected by attacks, with only a few false alarms. Scenario isolation was also precise, correctly identifying the nature of setting changes.
  • Robustness: Sensitivity analysis showed that the method maintains a degree of robustness to variations in controller settings and disturbance amplitudes.
  • Limitations: The method struggled to detect one specific attack scenario (ACT-2), where the controller could compensate for the attack, suggesting the need for incorporating process models for improved detection in such cases.

Limitations and Future Work

The authors acknowledge that detecting subtle changes in the presence of significant disturbances can be challenging. They also note that changes in system operating conditions might necessitate retraining models and re-tuning parameters.

Future work includes:

  • Developing an integrated approach that incorporates process and actuator models for enhanced detection and isolation.
  • Exploring trivalent evaluation methods beyond binary residual evaluation for more nuanced fault and cyber-attack identification.
  • Conducting experimental verification on a process simulator and, if possible, pilot implementations on real installations.

Summary updated

BibTeX
@inproceedings{sztyberbetley2023controller,
  author = {Anna Sztyber-Betley and Michal Syfert and Jan Maciej Kóscielny and Zuzanna Górecka},
  title = {{Controller Cyber-Attack Detection and Isolation}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=Ny2y23lpw9},
}
Figure