Identifying Vulnerabilities of Industrial Control Systems using Evolutionary Multiobjective Optimisation

Nilufer Tuptuk, Stephen Hailes

Researchers at University College London have developed a methodology to automatically discover critical vulnerabilities in Industrial Control Systems (ICS) using evolutionary algorithms. By framing cyber-attacks as a multi-objective optimization problem, the study demonstrates how combinatorial attacks can rapidly shut down infrastructure or cause severe economic damage while remaining undetected by modern machine learning defenses.

The Challenge of ICS Security

Industrial Control Systems—the operational technology behind power grids, water treatment, and chemical plants—have transitioned from physical isolation to networked environments. This shift has exposed them to sophisticated cyber threats like Stuxnet.

The primary difficulty in securing these systems lies in the combinatorial explosion of potential vulnerabilities. A typical plant contains hundreds of sensors and actuators. Identifying the specific combination of signals that, when manipulated via Denial of Service (DoS), integrity, or replay attacks, will cause catastrophic failure without triggering alarms is a search space far too large for manual engineering.

Methodology: Evolutionary Search for Vulnerabilities

The researchers utilized Evolutionary Multiobjective Optimisation (EMO) to automate this search. Instead of seeking a single optimal attack, they looked for a Pareto front—a set of strategies representing trade-offs between attack impact (e.g., shutdown speed or economic cost) and the "effort" required (the number of compromised signals).

Key Techniques:

  • EMO Algorithms: The study compared NSGA-II and SPEA2, finding that SPEA2 generally provided a more diverse and higher-quality set of attack strategies.
  • Attack Modeling: Four attack types were simulated: DoS (freezing signals), Integrity (forcing signals to min/max limits), and Replay (masking malicious activity with recorded normal data).
  • Adversarial Testing: Attacks were evolved specifically to bypass Intrusion Detection Systems (IDS) built on machine learning models like Random Forest.

The Tennessee Eastman Benchmark

To test their methodology, the researchers used the Tennessee Eastman (TE) process model, a high-fidelity industry-standard simulator of a chemical plant. This complex, non-linear system includes a reactor, condenser, separator, compressor, and stripper, featuring 41 process measurements (sensors) and 12 manipulated variables (actuators).

The Tennessee Eastman process model schematic, showing the interconnected sensors (y) and actuators (u) across the plant.

Key Results and Insights

The EMO approach significantly outperformed random search, identifying devastating attacks that were previously unknown.

1. Rapid Plant Shutdowns

The EMO identified combinations of 6-7 sensors/actuators capable of forcing a total plant shutdown in approximately 8.3 minutes (0.138 hours). This speed is achieved by exploiting cascading failures that standard safety systems cannot mitigate in time.

2. Economic Sabotage and Stealth

For "stealthy" attacks designed to evade detection, the EMO successfully increased hourly operating costs from a baseline of $8,208 to over $56,000. The algorithms discovered that certain sensors are "noisier," allowing attackers to hide malicious signal changes within natural operational variance.

3. Vulnerability Mapping

The study mapped specific high-value targets. For example, manipulating XMEAS 4 (A and C feed flow) and XMEAS 9 (Reactor temperature) proved highly efficient for triggering pressure-related shutdowns.

A detailed analysis of sensor (XMEAS) vulnerabilities, showing the maximum cost and shutdown frequency for various attack types.

As shown in the data above, certain variables like Reactor Pressure (XMEAS 7) and Reactor Temperature (XMEAS 9) are critical points of failure. While some attacks on these sensors cause immediate shutdowns, others are optimized to maximize economic cost without triggering a full system halt.

Conclusion

This research highlights a significant gap in current ICS defenses. While machine learning-based IDS can be effective against standard attacks, they are vulnerable to evolved, multi-signal strategies. The authors suggest that future work should focus on co-evolution, where defensive systems and attack strategies evolve simultaneously to stay ahead of increasingly sophisticated threats.

Summary updated

BibTeX
@article{tuptuk2024identifying,
  author = {Nilufer Tuptuk and Stephen Hailes},
  title = {{Identifying Vulnerabilities of Industrial Control Systems using Evolutionary Multiobjective Optimisation}},
  journal = {Comput. Secur.},
  year = {2024},
  doi = {10.1016/j.cose.2023.103593},
}
Figure