Cyber-Physical Systems (CPS)—ranging from power grids and railway networks to autonomous vehicles—rely on a fundamental feedback loop: sensors measure the physical state, and controllers issue commands to actuators to maintain stability. Traditionally, these systems were secured via "air-gapping" or standard IT encryption. However, as Roshan K. Thomas et al. argue, traditional IT security is insufficient because it fails to account for the physical dynamics of the system.
A sophisticated class of threats known as False Data Injection Attacks (FDIAs) exploits this gap. Instead of crashing a system, an attacker injects malicious data into sensor streams or control commands to deceive the controller into taking actions that lead to physical damage, economic loss, or instability.
The Mechanism of Stealth
The core challenge of FDIAs is their "stealthiness." Most industrial control systems employ Bad Data Detection (BDD) mechanisms that filter out statistical outliers or signals that violate the laws of physics. As Zhenyong Zhang et al. explain, attackers can bypass these filters by operating within a "stealthy attack space." By understanding the system's physical and statistical models, an adversary can craft perturbations that mimic legitimate system behavior or standard measurement noise.
This vulnerability is particularly acute in Data-Driven Predictive Control (DPC). Because these systems replace explicit physics-based models with historical data, Ruihan Zhao et al. demonstrate that they are highly fragile to "smart" adversarial data. In such cases, the controller treats external forecasts as ground truth, allowing an attacker to drive the system toward mathematically "optimal" but physically catastrophic states. For example, Po-han Li et al. show that tiny, human-imperceptible perturbations in electricity demand forecasts can increase control costs by over 8000%.
Active Defense via Dynamic Watermarking
To counter stealthy attacks, researchers have moved toward active detection strategies. A leading approach is Dynamic Watermarking (DW). The intuition behind DW is similar to a secret handshake: the controller injects a low-magnitude, random "watermark" signal into the control command. Because the controller knows this secret signal, it expects to see its effects reflected in the subsequent sensor measurements.
Tong Huang et al. demonstrate that this allows operators to mathematically verify the integrity of sensor data even against "omniscient" attackers. If an attacker intercepts and modifies the sensor data but fails to account for the secret watermark, the lack of correlation between the command and the feedback exposes the injection. Dajun Du et al. further show that by integrating watermarking with compensation mechanisms, systems can achieve robust recovery with zero performance loss under normal conditions.
Moving Target Defense and State Decomposition
Another strategy involves Moving Target Defense (MTD), which aims to invalidate the attacker's knowledge of the system. By dynamically altering system parameters (such as line impedances in a power grid), Zhenyong Zhang et al. argue that the "stealthy attack space" can be reduced, forcing the attacker to either remain inactive or risk detection.
However, detection is only half the battle. Effective security requires active compensation. The authors of a recent study on state-space decomposition propose decoupling sensor-based attacks from actuator-based attacks. This allows the system to identify the specific source of the corruption and adjust the control law in real-time to maintain physical stability, even while the attack is ongoing.
Cross-Domain Vulnerabilities
The threat of stealthy injection extends beyond traditional infrastructure:
- Transportation: Subhash Lakshminarayana et al. highlight how railway traction power systems can be tricked into wasting energy or tripping circuit breakers through false voltage readings.
- Automotive: Ishtiaq Rouf et al. found that tire pressure monitoring systems lack basic input validation, allowing for forged packets that override legitimate sensor data.
- Environmental: In freshwater monitoring, the authors of a cyberbiosecurity study argue that manipulating IoT sensor data can mask toxic algal blooms, preventing the activation of critical water filtration protocols.
Ultimately, as Hamza Fawzi et al. conclude, securing these systems requires a "separation principle": combining resilient state estimators that can handle adversarial, unbounded signals with standard control laws to ensure that the physical reality remains aligned with the digital perception.