Physics-Constrained Adversarial Machine Learning

23 concepts15 papers

In traditional computer vision, generating an adversarial attack is mathematically unconstrained. An attacker can manipulate individual pixels arbitrarily to trick a neural network into misclassifying an image. Because any combination of pixel values represents a mathematically valid image, the digital space offers infinite freedom to the adversary.

However, when machine learning models are deployed to manage critical infrastructure—such as power grids, water distribution networks, autonomous vehicles, or medical imaging systems—this unconstrained assumption completely breaks down. Physical systems are governed by immutable laws: conservation of energy, fluid dynamics, kinematics, and network topologies. If an attacker injects arbitrary digital noise into a real-world sensor stream, legacy "bad data detection" (BDD) systems or hardware sanity checks will instantly flag the data as physically impossible.

To successfully deceive machine learning models operating within cyber-physical systems, adversarial perturbations must be explicitly constrained by the laws of physics.


The Core Mechanism: Shifting from Sensor Space to State Space

To understand why physics-constrained attacks work, we must look at how they are generated. If an adversary attempts to manipulate raw sensor measurements directly, balancing the need to evade an AI model while simultaneously satisfying complex physical equations is an incredibly difficult optimization problem.

To overcome this, researchers have shifted the attack generation process from the sensor measurement space to the underlying physical state space.

Unconstrained Attack (Fails):
[Adversarial Optimization] ──> Perturbed Sensors (y + ε) ──> Caught by Physical Filters (BDD)

Physics-Constrained Attack (Succeeds):
[Adversarial Optimization] ──> Perturbed Physical States (x + ε) ──> [Forward Physical Model] ──> Plausible Sensors (y_adv) ──> Bypasses Filters

As demonstrated by Jiwei Tian et al. (2024) and Jiwei Tian et al. (2026) in the context of smart grids, an adversary can optimize a perturbation directly within the system's underlying state variables (such as voltage angles or nodal pressures) rather than the raw sensor readings. Because the final manipulated sensor readings are derived by passing these altered states through the system's true topological and physical equations, the resulting adversarial data inherently satisfies domain-specific physical laws.

Similarly, Obai Bahwal et al. (2024) show that attackers can apply inverse transformations to perturbed features. By reconstructing time-series signals directly from manipulated physical characteristics, the resulting malicious data streams appear entirely legitimate to automated classifiers while successfully blinding the underlying models.


Mapping Constraints Across Physical Domains

The specific mathematical constraints required to maintain stealth vary depending on the physics of the domain:

Power Grids and Smart Infrastructure

Modern grids increasingly rely on machine learning for load forecasting and event identification, yet these models often lack an internal understanding of physical laws (Yize Chen et al., 2018). To bypass legacy monitoring tools, adversarial data must strictly adhere to network topology and Kirchhoff's Laws (Zhenyong Zhang et al., 2023). Without modeling these intricate dependencies, generated attacks lack the realism required to escape detection by physical-layer anomaly detectors (Xu Wan et al., 2022).

Fluid Dynamics and Water Networks

In water distribution networks, any simulated anomaly or leak must remain hydraulically consistent to avoid detection (André Artelt et al., 2023). An attacker seeking to maximize physical damage without triggering alarms must ensure that the simulated pressure drops and flow rates across a pipe network perfectly match the non-linear equations of fluid mechanics.

Kinematics and Trajectory Tracking

For autonomous vehicles and drones, erratic or discontinuous data signals immediate sensor failure. To remain stealthy, trajectory perturbations must respect kinematic constraints like acceleration limits and smooth motion profiles (Yiming Li et al., 2021; Kaiyuan Tan et al., 2023). Furthermore, the authors of Stealthy Perception-based Attacks on Unmanned Aerial Vehicles demonstrate that to maintain stealth within a real-world control loop, an attacker must simultaneously falsify camera images and physical sensor data (like GPS and IMU) so that both streams remain mathematically consistent with the vehicle's flight dynamics.

Wave Mechanics and Atmospheric Physics

This paradigm extends to advanced sensing and environmental modeling. In Synthetic Aperture Radar (SAR) target recognition, Bowen Peng et al. (2022) generate adversarial examples by optimizing the parameters of a physical scattering model rather than individual pixels. This ensures the perturbation mimics real electromagnetic reflections from geometric structures. In deep learning-based weather forecasting, adversarial modifications must respect the complex, non-linear laws of atmospheric physics to avoid being flagged by quality control systems or human meteorologists (Francisco Santos et al., 2025).


Physics as a Defensive Guardrail

While physical laws complicate the generation of adversarial attacks, they also provide defenders with a powerful, immutable tool for resilience. Standard data-driven defenses developed for computer vision, such as image filtering, often fail in cyber-physical systems because they inadvertently alter data characteristics and degrade the physical meaning of time-series sensor data (Yize Chen et al., 2018).

Instead, effective defenses embed physical invariants directly into the model training or inference loops:

  • Temporal and Dynamic Filtering: By moving beyond static "snapshot" checks and modeling system physics over time, defenders can expose coordinated, multivariate attacks that appear mathematically plausible at any single instant but violate dynamic conservation principles (Kaikai Pan et al., 2019; Hassan Ali et al., 2023).
  • Inference-Time Physics Scrubbing: In computational MRI reconstruction, Mahdi Saberi et al. (2026) introduce a training-free defense that leverages cyclic measurement consistency—a known physical forward model—to scrub away adversarial noise during inference without needing to retrain the underlying neural network.

Ultimately, physics-constrained adversarial machine learning demonstrates that security in critical infrastructure does not have to rely on the secrecy of the detection mechanism. If the system's physical constraints are sufficiently tight, optimal detection performance can be achieved because the laws of nature act as the ultimate validation metric.

Key Foundational Work

Is Machine Learning in Power Systems Vulnerable? (2018) — contributed 4 concept elements to this direction.
From Static to Dynamic Anomaly Detection With Application to Power System Cyber Security (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physics_constrained_adversarial_perturba_c99c9a9d,
  title = {physics-constrained adversarial perturbation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-constrained-adversarial-perturbation/rdr_c99c9a9de51c47598952445387afc83b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_physics_constrained_adversarial_perturba_c99c9a9d_v1,
  title = {physics-constrained adversarial perturbation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-constrained-adversarial-perturbation/rdr_c99c9a9de51c47598952445387afc83b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}

See something wrong? Report a mistake