In modern engineering, we rarely build systems that are "stable" on their own. Instead, we build systems that are kept stable by constant, aggressive correction. A modern fighter jet, a high-voltage power converter, or an anti-lock braking system (ABS) are often "open-loop unstable"—without a computer constantly adjusting their state based on sensor data, they would quickly spin out of control or fail.
This reliance on feedback creates a profound security paradox. The very mechanism designed to ensure safety—the feedback loop—can be weaponized to drive a system into a catastrophic state. By injecting false data into sensors, an attacker doesn't just break the system; they trick the system into breaking itself.
The Feedback Loop as a Lever
The core intuition behind this research direction is that a controller is a reactive machine. If a sensor reports that a value is too low, the controller will automatically increase it. Marcell Szakály, Martin Strohmeier, and their colleagues demonstrated this in Assault and Battery: Evaluating the Security of Power Conversion Systems Against Electromagnetic Injection Attacks. By using electromagnetic interference to spoof a battery's voltage sensor, they could trick a charger into "thinking" a battery was empty when it was actually full. The charger’s own feedback logic then dutifully overcharged the battery, leading to hardware destruction.
This is not a simple "off" switch. It is a "leverage" attack. As Paul D. Martin 0001, Yasser Shoukry, and others showed in Non-invasive Spoofing Attacks for Anti-lock Braking Systems, spoofing just one wheel's speed sensor during braking causes the ABS to create a torque imbalance. The car’s own safety system effectively steers the vehicle off the road.
The Paradox of Aggressive Stability
One might assume that a more robust, highly stable controller would be harder to attack. However, research by Amir Khazraei and Miroslav Pajic suggests the opposite. In Attacks on Perception-Based Control Systems: Modeling and Fundamental Limits, they argue that aggressive feedback mechanisms—those designed to return a system to equilibrium as fast as possible—actually expand the "attack surface."
Because these controllers react so strongly to any perceived deviation, they follow falsified signals more effectively. An attacker can use this high "gain" to steer the system into a dangerous state while keeping the injected signals subtle enough to bypass anomaly detectors. This creates a fundamental tradeoff: the more a controller is tuned for high performance and tight regulation, the more capacity it provides for an attacker to inject malicious physical deviations.
Stealth Through System Dynamics
The most sophisticated versions of these attacks are "stealthy," meaning they are mathematically invisible to standard monitors. In Vulnerability Analysis of Nonlinear Control Systems to Stealthy False Data Injection Attacks, Khazraei and Pajic show that if an attacker understands the physical model of the system, they can inject data that perfectly mimics a valid (but incorrect) environmental state.
The controller then "corrects" for this fake state. Because the physical plant is often unstable in its open-loop state, the controller’s corrective actions drive the actual physical state away from the target. To the system's internal monitors, everything looks perfect—the sensors are reporting exactly what the controller expects to see after its "correction"—while the physical reality is drifting toward failure.
Moving Toward Active Defense
If the problem is that the controller blindly trusts its feedback, the solution must involve verifying that feedback. Tong Huang, Bharadwaj Satchidanandan, and Le Xie propose a technique called Dynamic Watermarking in An Online Detection Framework for Cyber Attacks on Automatic Generation Control. By superimposing a secret, low-magnitude random signal onto the control commands, the operator can check if that same "watermark" appears in the returning sensor data. If the sensor data doesn't reflect the watermark, it has been tampered with.
Other researchers, such as Yuzhou Xiao and Yuanqing Xia, suggest that in the face of severe attacks, the safest move is to stop following the feedback entirely. In Stochastic Tube-based Model Predictive Control for Cyber-Physical Systems under False Data Injection Attacks with Bounded Probability, they demonstrate that switching to a pre-calculated "open-loop" mode can be safer than following corrupted sensor data, which often destabilizes a system faster than passive drift.