Abstract
In this paper, we investigate the detectability and identifiability of attacks on linear dynamical systems subjected to external disturbances. We generalize the concept of a security index, previously introduced for static systems, to a dynamical context. This generalized index exactly quantifies the resources necessary for targeted attacks to remain undetectable and unidentifiable in the presence of disturbances. This information is critical for both risk assessment and the design of robust anomaly detectors. Finally, we demonstrate how techniques from fault detection literature can be used to decouple disturbances and identify attacks under specific sparsity constraints.
Introduction
Modern industrial control systems—ranging from power grids and water treatment plants to automated manufacturing lines—increasingly rely on networked communication. While this connectivity improves efficiency, it introduces significant vulnerabilities to cyber-physical attacks. A sophisticated adversary does not merely inject noise; they craft false data designed to mimic natural physical disturbances or routine sensor faults. If an attack can be perfectly explained away by a plausible natural disturbance, the system operator remains blind to the intrusion.
This research addresses the fundamental mathematical limits of such deception. By quantifying the resources an attacker needs to remain hidden, we establish theoretical bounds on when a defender can reliably detect and uniquely reconstruct (identify) an attack despite background noise and system disturbances.
The Dynamical Security Index
To analyze these vulnerabilities, the study models the environment as a discrete-time linear dynamical system. The core contribution is the introduction of the Dynamical Security Index ($\alpha_i$).
This index is formulated as a combinatorial optimization problem that calculates the sparsest possible attack—the minimum number of compromised signals—required to perfectly cancel out the attack's footprint in the system's output. By hiding behind a valid system disturbance and initial state, the attack becomes "stealthy."
To evaluate these "blind spots," the authors utilize the Rosenbrock System Matrix pencil. This mathematical tool allows for the identification of specific frequencies or input patterns that result in a zero-output response, effectively mapping the system's dynamical vulnerabilities.
The "Half-Distance" Rule for Identifiability
A pivotal insight of this work is the mathematical relationship between the difficulty of detecting an attack and the difficulty of identifying it. This relationship mirrors principles found in coding theory and compressed sensing.
- Undetectability Bound: A persistent, undetectable attack against a specific system element $i$ is only possible if the attacker can simultaneously manipulate $q$ elements, where $q \geq \alpha_i$.
- Identifiability Bound: All persistent attacks are uniquely identifiable—meaning the exact attack vector can be reconstructed—if and only if the number of compromised elements $q$ is less than half the minimum security index: $q < \min_i \alpha_i / 2$.
This "half-distance" rule provides a clear threshold for security: if a defender knows the security index of their system, they can determine exactly how many sensors or actuators must be hardened to guarantee that any successful breach can be mathematically isolated.
Decoupling and Reconstruction
To move from theory to defense, the authors adapt techniques from classical Fault Detection and Isolation (FDI). While traditional FDI assumes anomalies are independent and non-malicious, cybersecurity must account for colluding and intentional deception.
The paper proposes the use of left-invertible post-filters. These filters are designed to mathematically decouple natural disturbances from malicious attacks. By stripping away the "mask" of natural disturbances, the filter creates a residual signal that is influenced solely by the malicious input. Once the disturbances are removed, the problem of identifying the sparse attack signal becomes an inversion problem, similar to recovering a signal in compressed sensing (where the security index acts as the "spark" of the system matrix).
Limitations and Future Outlook
While the theoretical framework is robust, practical implementation faces two primary hurdles:
- Computational Complexity: Calculating the security index $\alpha_i$ involves solving an $L_0$-norm minimization problem, which is NP-hard. While polynomial-time solutions exist for specific structures (like certain power grid topologies), computing this for arbitrary large-scale systems remains a challenge.
- Combinatorial Explosion: Reconstructing the attack signal requires evaluating numerous combinations of potentially compromised nodes. As the system size and the number of possible compromised elements $q$ increase, the computational cost of this inversion grows significantly.
Future research is directed toward refining the algorithmic efficiency of the identification procedure and developing methods to better handle the transients of unknown initial states in real-world, noisy environments.
Summary updated