Traditional security for industrial plants and power grids often relies on "static" state estimation. These systems look at a single snapshot of data—sensor readings, voltages, or temperatures—and ask: Does this look valid right now? If the values fall within expected bounds and satisfy basic algebraic equations (like Kirchhoff’s laws), the system is deemed secure.
However, research across cyber-physical systems (CPS) has revealed a fundamental flaw in this approach. Sophisticated attackers can craft "stealthy" False Data Injection (FDI) attacks that are mathematically invisible to static detectors. By coordinating corrupted measurements across multiple sensors, an adversary can trick a controller into believing the system is in a safe state when it is actually being pushed toward failure.
The solution emerging from recent literature is a shift toward dynamic monitoring. Instead of analyzing isolated snapshots, dynamic monitors leverage the physical laws governing how a system evolves over time.
The Limitation of the Static Snapshot
In a static framework, a monitor only checks for spatial consistency. As Fabio Pasqualetti demonstrates in Attack Detection and Identification in Cyber-Physical Systems, these monitors are fundamentally limited to detecting simple sensor-level discrepancies. They cannot identify attacks that manipulate the underlying physical dynamics of the system.
Consider a power grid. A static monitor might see a set of bus voltages that perfectly satisfy the power flow equations. But as Ali Sayghe and colleagues point out in their survey of machine learning methods, these snapshots ignore the "temporal correlations" of the grid. A measurement might be mathematically plausible in a vacuum, yet completely impossible given the system's state five seconds ago.
Leveraging Physical Inertia
The core intuition behind dynamic monitoring is that physical systems have inertia. A massive turbine cannot change its rotation speed instantly; a large chemical tank cannot empty in a millisecond.
Kaikai Pan, Peter Pálenský, and Peyman Mohajerin Esfahani formalize this in From Static to Dynamic Anomaly Detection. They use "dynamic residual generation"—essentially a digital filter that compares incoming measurement streams against a model of the system's physics. If the sequence of measurements violates the system's temporal trajectory or control loop signatures, the residual deviates from zero, exposing the attack. Even if an attacker knows the static equations of the grid, they must also mimic the complex, time-dependent behavior of the entire physical process to remain hidden.
Detecting Stealth through Temporal Patterns
When we move from static to dynamic analysis, the "stealthiness" of an attack becomes a trade-off. In Balancing detectability and performance of attacks, Alessio Russo and Alexandre Proutiere show that while an attacker can minimize the magnitude of individual perturbations to avoid simple triggers, these small changes accumulate over time into statistically significant deviations that dynamic monitors can catch.
Modern deep learning architectures are particularly well-suited for this task. For instance, Wangkun Xu and Fei Teng propose a Long Short-Term Memory (LSTM) Autoencoder that learns the normal temporal patterns of a power system. By analyzing data across a sliding temporal window, the model identifies inconsistencies in the "flow" of measurements that static analysis would overlook.
This temporal approach also solves practical industrial problems beyond security. In Distributed dynamic modeling and monitoring, Chunhui Zhao uses Sparse Slow Feature Analysis to distinguish between real faults and routine changes in operating conditions. By focusing on the speed and nature of variable changes rather than just their current values, the system reduces "nuisance alarms" that plague static monitoring.
The Statistical Shift
Finally, dynamic monitoring changes how we define "normal." Instead of checking if a single data point is an outlier, researchers like Christoph H. Lampert suggest monitoring the distribution of data over time. Their KS(conf) test applies statistical tests to batches of data to detect distribution shifts. This can catch "static failures"—like a frozen sensor that keeps reporting the same "valid" image—which a snapshot-based monitor would never flag as suspicious because each individual frame looks perfect.
By treating cyber-physical security as a problem of trajectory verification rather than state estimation, these researchers have provided a more robust defense against adversaries who know the system's rules but cannot easily replicate its physical reality over time.