In traditional cybersecurity, an attack is often a binary event: a password is stolen, a firewall is breached, or a service is taken offline. However, in Cyber-Physical Systems (CPS)—the networks that control power grids, autonomous vehicles, and industrial robots—a more insidious class of threat exists. These are "stealthy" attacks, where an adversary manipulates sensor or actuator data so subtly that the system continues to appear healthy while being driven toward a catastrophic state.
The Illusion of Normalcy
The core of a stealthy attack is the exploitation of the gap between physical reality and digital reporting. Most modern control systems use anomaly detectors that monitor "residues"—the difference between what a sensor reports and what the system's mathematical model predicts should happen. If the residue exceeds a certain threshold, an alarm triggers.
As André M. H. Teixeira, Iman Shames, and Henrik Sandberg demonstrate in A Secure Control Framework for Resource-Limited Adversaries, an attacker with knowledge of the system's model can craft signals that stay perfectly within these thresholds. By injecting "zero-dynamics" or "bias injection" attacks, they ensure the residues remain zero or negligible, making the attack mathematically invisible to standard monitors.
Hiding in the Noise
Every physical sensor has inherent noise. A GPS receiver might fluctuate by a few centimeters; a thermometer might jitter by a fraction of a degree. Stealthy attacks, such as those explored by Tianci Yang and Chen Lv in Risk Assessment for Connected Vehicles under Stealthy Attacks on Vehicle-to-Vehicle Networks, hide malicious signals within this statistical uncertainty.
This creates a dangerous trade-off: if a defender makes the anomaly detector too sensitive, they suffer from constant false alarms due to natural noise. If they make it too lenient, the attacker has a "safe zone" to operate. This "hidden in plain sight" phenomenon isn't limited to physics; Nihar B Shah and Steven Jecmen found a similar "undetectable region" in academic peer review, where collusion rings can manipulate paper assignments while maintaining a statistical profile that appears normal.
Temporal Stealth and Bias Injection
One of the most effective ways to remain stealthy is to avoid sudden changes. A sharp spike in a sensor reading is easy to catch. Instead, attackers use temporal smoothing. In A Secure Control Framework for Resource-Limited Adversaries, the authors describe passing attack signals through a low-pass filter. This gradually ramps up a bias, moving the system to an unsafe state over a long duration.
In the context of transportation, Feilong Wang and Xuegang Ban show that stealthy GPS spoofing uses gradual drift. By slowly shifting a vehicle's perceived position, the attack bypasses internal consistency checks that would flag a sudden jump, eventually leading the vehicle off-course without ever triggering a "glitch" alert.
The Reachability of Damage
If an attack is so small it looks like noise, can it really do damage? The research suggests the answer is a resounding yes. Through reachability analysis, researchers like Zhicong Sun and Shuanghua Yang have shown that even sub-threshold perturbations can accumulate.
For example, in Vulnerabilities of Power System Operations to Load Forecasting Data Injection Attacks, Yize Chen demonstrates that a tiny, 1°F shift in weather forecast data—well within the range of normal error—can trick grid operators into making dispatch decisions that lead to massive financial losses or blackouts. Similarly, Siew Kei Lam notes that automated insulin pumps can be tricked by "process-aware" attacks that stay within physiological ranges but cause cumulative harm over time.
Defending the Physical Layer
Because these attacks exploit the system's own logic and noise, traditional digital security (like encryption) is insufficient. Defense must become "physics-aware."
- Side Information: Yuan Chen and José M. F. Moura propose using side information about the system's initial state—data an attacker cannot manipulate—to expose inconsistencies that would otherwise be hidden.
- Physical Watermarking: This involves adding a known, low-level "probe" signal to the control inputs. If the sensor returns don't reflect this secret watermark, the system knows the data has been tampered with. André M. H. Teixeira and colleagues explore this in Quickest Detection of Deception Attacks.
- Moving Target Defense: By constantly changing the detection thresholds or controller parameters, defenders can prevent an attacker from ever "learning" the model well enough to stay stealthy. David Umsonst and Henrik Sandberg formalize this as a Bayesian Nash equilibrium game, where the system's defensive posture is always shifting.
Ultimately, the shift in research direction from Husheng Li's early work on GPS spoofing in 2012 to modern distributed adaptive control highlights a fundamental truth: in the physical world, security is not just about keeping people out, but about ensuring that what we see is what is actually happening.