Abstract
This paper studies the physical consequences of unobservable false data injection (FDI) attacks designed only with information inside a sub-network of the power system. The goal of this attack is to overload a chosen target line without being detected via measurements. To overcome the limited information, a multiple linear regression (MLR) model is developed to learn the relationship between the external network and the attack sub-network from historical data. The worst possible consequences of such FDI attacks are evaluated by solving a bi-level optimization problem wherein the first level models the limited attack resources, while the second level formulates the system response to such attacks via DC optimal power flow (OPF). The vulnerability of this attack model is illustrated on the IEEE 24-bus RTS and IEEE 118-bus systems.
Introduction
Most research on False Data Injection (FDI) attacks assumes the adversary has "perfect knowledge"—a complete, god-level view of the entire power grid's topology and parameters. In reality, gaining such comprehensive access is incredibly difficult. This paper addresses a more realistic and chilling scenario: the limited information attacker.
As shown in Figure 1, the focus is on a specific subset of cyber threats: unobservable FDI attacks that lead to severe physical consequences despite the attacker having restricted access to system information. The challenge for the attacker is predicting how the rest of the grid (the "external" network) will react when they start spoofing data. If the attacker cannot predict the grid's re-dispatch response, their attack might fail to cause an overflow or, worse, get caught by bad data detection (BDD) systems.
Methodology
The researchers propose a sophisticated two-step strategy for the limited-information attacker to bridge the knowledge gap:
1. Learning the "Black Box"
The attacker treats the external network as a black box. By observing historical data (loads, generation, and prices) within their sub-network, they use Multiple Linear Regression (MLR) to estimate "pseudo-boundary injections." Essentially, they learn a linear relationship that predicts how power flows in and out of their sub-network based on local conditions.
Figure 2 illustrates this partitioning. The attacker possesses perfect knowledge only within the sub-network $\mathcal{L}$, while the area $\mathcal{E}$ remains unknown. The boundary buses serve as the interface where the regression model estimates the impact of the external network.
2. Bi-level Optimization
The attack is formulated as a mathematical "chess match":
- Upper Level: The attacker chooses an attack vector to maximize power flow on a target line while staying within a "budget" (changing as few measurements as possible).
- Lower Level: This models the grid operator’s response. The attacker uses a modified DC Optimal Power Flow (OPF) that incorporates their learned MLR model to predict how the control center will re-dispatch generation after the data is injected.
To make this computationally feasible, the authors convert this complex bi-level problem into a Mixed-Integer Linear Program (MILP) using Karush-Kuhn-Tucker (KKT) conditions.
Key Insights and Results
The study reveals several critical findings regarding grid vulnerability:
- The Rank Condition (Theorem 1): The authors prove that the attacker’s ability to perfectly predict the external network depends on the rank of a specific matrix related to congested lines. If the number of congested lines is sufficient relative to the number of buses, the "black box" becomes perfectly transparent to the regression model.
- Stealth via Localized Spoofing: By only modifying measurements within their sub-network and ensuring the "spoofed" state looks like a valid power flow solution, the attacker bypasses the grid's Bad Data Detection (BDD) entirely.
- Physical Consequences: In both the 24-bus and 118-bus systems, the limited-information attacker successfully caused target lines to exceed their thermal limits. While the "local" attack is less "efficient" than a "global" attack (requiring more resources), it still achieves the primary goal of physical disruption.
Conclusion
This research demonstrates that an attacker does not need a full map of the power grid to cause significant damage. By exploiting historical data, they can effectively "see" through the boundaries of their limited access. While topology changes or a lack of historical data can hinder such attacks, the robustness of the results against non-linear AC models suggests that current power systems face a significant threat from localized, data-driven FDI attacks. Future work will explore defensive strategies, such as using machine learning to detect these subtle, localized patterns of data injection before they lead to physical failure.
Summary updated