Centrality-Based Security Allocation in Networked Control Systems

Anh Tung Nguyen, Andreas Hertzberg, André M. H. Teixeira

TL;DR

This paper tackles the critical problem of allocating security resources in networked control systems (NCS) that are vulnerable to stealthy cyberattacks. The authors propose a novel centrality-based approach to efficiently identify the most influential nodes to monitor, significantly reducing computational costs compared to optimal solutions while maintaining acceptable performance.

What Problem Does This Paper Address?

Modern infrastructure, like power grids and transportation networks, relies heavily on interconnected networked control systems (NCS). These systems, while efficient, are increasingly vulnerable to cyberattacks due to their reliance on open communication technologies. Stealthy data injection attacks, in particular, pose a significant threat as they can disrupt system performance without immediate detection. The challenge lies in effectively allocating limited security resources (e.g., sensors for monitoring) to detect and mitigate these attacks. Traditional approaches often involve computationally intensive methods that consider all possible attack scenarios, making them impractical for large-scale systems. This paper addresses this gap by proposing a more efficient method for security resource allocation.

Methods

The core of this paper is a centrality-based security allocation methodology. The authors leverage graph theory concepts to identify critical nodes within the NCS network. Specifically, they propose using three common centrality measures:

  • Degree Centrality: Identifies nodes with the most direct connections.
  • Closeness Centrality: Identifies nodes that are, on average, closest to all other nodes.
  • Betweenness Centrality: Identifies nodes that lie on the most shortest paths between other nodes.

The approach involves:

  1. Generating all possible monitor sets of a given size (sensor budget).
  2. Calculating the total centrality for each monitor set based on the chosen centrality measure.
  3. Selecting the monitor set with the highest total centrality.
  4. For cases with multiple sets having the same highest centrality, they evaluate each and select the one with the lowest worst-case attack impact (WCAI).
  5. A combined centrality measure is also explored, which considers the best performing set across all three individual measures.

This centrality-based selection is then used to inform the security allocation, aiming to minimize the WCAI. The authors compare this approach against an optimal solution derived from a mixed-integer SDP problem.

Datasets

The paper evaluates its proposed method using two types of network structures:

  • Erdős-Rényi random graphs: These are generated with a probability of 0.5 for edge inclusion, simulating various network topologies. Experiments are conducted on networks of sizes N = {10, 12, 14, 16, 18, 20} with attack budgets na = {1, 2} and monitor budgets ns = 1.
  • IEEE 14-bus system: This is a standard benchmark for power systems, providing a more realistic and complex network structure for evaluation.

Key Insights and Ideas

  • Centrality as a proxy for vulnerability/importance: The fundamental idea is that nodes with high centrality are likely to be more influential in the network's overall performance and thus more critical targets for adversaries or more effective for defender monitoring.
  • Computational efficiency is paramount: The paper strongly emphasizes that traditional optimal solutions are computationally prohibitive for large networks. The centrality-based approach offers a significant speedup, making security allocation feasible in practice.
  • Combined centrality offers a robust solution: While individual centrality measures show promise, a combined approach (taking the best of degree, betweenness, and closeness) often yields superior results, balancing different aspects of network influence.
  • Betweenness centrality stands out: Among the individual centrality measures, betweenness centrality consistently performed well in terms of minimizing WCAI and reducing computation time.

Key Results

  • Reduced WCAI gap: The centrality-based approaches, particularly the combined and betweenness centrality measures, achieved a relatively small gap in worst-case attack impact (WCAI) compared to the optimal solution, often staying under 10% for Erdős-Rényi graphs.
  • Significant speedup in computation: The proposed centrality-based methods drastically reduced the solving time compared to the optimal solution. For instance, the combined centrality approach reduced solving time by around 70%, and betweenness centrality alone by about a third or more.
  • Scalability: The method demonstrates effectiveness on networks up to 20 nodes and shows promise for larger systems where optimal solutions are infeasible.
  • IEEE 14-bus system validation: On the IEEE 14-bus system, the betweenness centrality measure identified a monitor vertex (bus 4) that resulted in a WCAI gap of under 2% compared to the optimal solution (bus 2), highlighting its practical applicability.

Limitations and Future Work

  • Alarm threshold is fixed: The paper assumes the alarm threshold ($\delta$) is given and fixed, not optimized by the defender. Future work could explore optimizing this threshold alongside monitor selection.
  • Focus on stealthy FDI attacks: The current work specifically addresses stealthy false data injection (FDI) attacks. Extending the methodology to other types of cyberattacks would be valuable.
  • Limited monitor budget: The experiments primarily focus on a monitor budget of $n_s=1$. Investigating the performance with larger sensor budgets could reveal further insights.
  • Erdős-Rényi vs. real-world networks: While Erdős-Rényi graphs provide a good baseline, evaluating the method on more realistic network models or actual system data would strengthen its validation.
  • Adversary's knowledge: The paper assumes the adversary has some knowledge of the system but doesn't explicitly model adaptive adversaries who learn and change strategies based on the defender's actions. Future research could explore more sophisticated adversary models.

Summary updated

BibTeX
@inproceedings{nguyen2024centrality,
  author = {Anh Tung Nguyen and Andreas Hertzberg and André M. H. Teixeira},
  title = {{Centrality-Based Security Allocation in Networked Control Systems}},
  booktitle = {Critis},
  year = {2024},
  doi = {10.1007/978-3-031-84260-3\_13},
}
Figure