Stealthy Adversarial Attacks via Sensor Perception Gaps

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 219 papers36 conceptsfrozen 2026-10-04

In the classic framing of adversarial machine learning, an attacker adds "noise" to an image to trick a model. In the digital world, this noise is mathematically constrained to be nearly invisible to humans. However, in the physical world, creating a "stealthy" attack is much harder. A giant sticker on a stop sign or a strange pattern on a face is immediately obvious to a human driver or security guard.

To solve this, researchers have moved beyond trying to hide noise in plain sight. Instead, they exploit the fundamental hardware discrepancies between how human eyes and machine sensors perceive the world. By operating in the "blind spots" of human biology—such as the infrared spectrum, light polarization, or high-frequency flickering—attackers can inject malicious data that is vivid to an AI but completely invisible to a person.

The Spectral Sensitivity Gap

The most common vector for these attacks is the difference in spectral range. While humans are limited to the visible light spectrum (roughly 400–700nm), the CMOS and CCD sensors used in cameras are naturally sensitive to near-infrared (NIR) and ultraviolet (UV) light.

In Invisible Mask: Practical Attacks on Face Recognition with Infrared, Kehuan Zhang and Weili Han demonstrate that 850nm infrared light is captured by cameras as visible color but remains imperceptible to humans. By hiding IR LEDs in a baseball cap, an attacker can project adversarial "spots" onto their face. To a security camera, the person’s digital embedding is completely altered, yet to a bystander, they are just wearing a normal hat.

This vulnerability is exacerbated by hardware design choices in autonomous systems. As Takami Sato and Sara Rampazzi point out in their work on Infrared Laser Reflection Attacks, many autonomous vehicles omit infrared-cut filters to improve low-light sensitivity. This creates a massive security hole: an attacker can use an IR laser to reflect invisible patterns off a "Stop" sign, causing the car to misclassify it as a "Yield" sign with 100% success, while the driver sees nothing unusual.

Exploiting Polarization Blindness

Light has properties beyond just color and intensity, such as polarization (the orientation of light waves). While some animals can see polarization, humans are naturally "polarization-blind." Computer vision systems, however, increasingly use polarization sensors to see through glass or map 3D surfaces.

In Fooling Polarization-Based Vision Using Locally Controllable Polarizing Projection, Zhuoxiao Li and Yinqiang Zheng show that this creates a perfect channel for stealth. By modifying a standard projector to manipulate polarization rather than color, they can project high-magnitude adversarial perturbations. Because the human eye only integrates the total intensity of the light, the scene looks perfectly normal to us, but the polarization-aware AI sees a completely corrupted input.

Latent Attacks via Fluorescent Conversion

A significant challenge for infrared-based attacks is that they can be defeated by simple hardware filters. To bypass this, researchers like Shuai Yuan and Hongwei Li have developed "latent" attacks using fluorescent materials.

In The Fluorescent Veil and ITPatch, they describe using invisible fluorescent ink that only emits visible light when triggered by a specific ultraviolet (UV) source. This allows an attacker to apply a seemingly transparent patch to a traffic sign. The attack remains dormant and invisible to everyone until the attacker (or the vehicle's own headlights, if they contain UV components) triggers the ink, which then emits a visible-spectrum adversarial pattern that passes right through the camera's filters.

Cross-Modal Stealth and Software Gaps

The perception gap isn't limited to light frequency; it also exists in how software processes data. In Can You Trust What You See? Alpha Channel No-Box Attacks on Video Object Detection, Ariana Yi and Qiben Yan exploit the "alpha channel" (transparency) in video files. Human-facing video players render the alpha channel to blend images, but many AI preprocessing pipelines simply discard it. By hiding a malicious scene in the RGB channels and using the alpha channel to make it transparent to the human viewer, they ensure the AI sees a "malicious" frame while the human sees a "benign" one.

Similarly, in the thermal domain, Hui Wei and Yinqiang Zheng developed the HOTCOLD Block. This attack uses materials that are opaque to visible light (looking like normal clothing) but have specific thermal signatures. By arranging these materials, they can make a person "invisible" to thermal detectors used in surveillance, achieving stealth across different sensing modalities.

Why This Works: The Asymmetry of Perception

The core intuition behind this research direction is asymmetry. Security systems often rely on a "human-in-the-loop" to verify what the AI is seeing. However, if the attacker can find a physical property that the sensor measures but the human eye ignores, they can decouple the two observers.

As Junbin Fang notes in his survey of optical attacks, these hardware-level differences—from CMOS spectral sensitivity to the way satellite sensors like those studied by Andrew Du use multispectral bands—create a unique attack surface that cannot be patched by software alone. To defend against these, we must align the "vision" of the machine more closely with the biological constraints of the human.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 9 papers
  2. Version 1 · 2026-10-04 · 19 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_cross_modal_perception_gap_6933978e,
  title = {cross-modal perception gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-modal-perception-gap/rdr_6933978e78604d8385969353a4babc9d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_stealthy_adversarial_attacks_via_sensor__6933978e_v1,
  title = {Stealthy Adversarial Attacks via Sensor Perception Gaps},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stealthy-adversarial-attacks-via-sensor-perception-gaps/rdr_6933978e78604d8385969353a4babc9d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}