Adversarial Attacks Using Dynamic Light Projection

This direction is merged into another direction: physical-world adversarial robustness, dynamic adversarial projection. Its last version stays readable and citable here.

Version 1 of 122 papers30 conceptsfrozen 2026-10-04

Traditional physical adversarial attacks rely on static modifications to the environment, such as placing stickers on stop signs or wearing 3D-printed glasses. While effective in controlled settings, these methods are rigid: they cannot adapt to changing lighting, they are easily spotted by human observers, and they leave permanent physical evidence.

A new frontier of research shifts the medium of attack from matter to light. By using projectors, lasers, and high-frequency LEDs, researchers can manipulate the raw signals entering a camera sensor in real-time. These optical attacks are transient, reconfigurable, and often invisible to the human eye, posing a significant challenge to the "eyes" of autonomous systems.

The Shift from Static to Dynamic Perturbations

The core intuition behind light-based attacks is that computer vision models do not see "objects"—they see patterns of photons. If an attacker can control those photons before they reach the lens, they can rewrite the digital representation of an object without ever touching it.

Giulio Lovisotto and colleagues demonstrated this with SLAP (Short-Lived Adversarial Perturbations). Instead of a permanent sticker, SLAP uses a projector to cast light patterns onto a stop sign. This allows the attack to be toggled on or off instantly. Because the perturbation is light-based, it can bypass "patch-detection" defenses that look for physical anomalies like tape or stickers. Similarly, Dinh-Luan Nguyen showed that light projection can be used against face recognition systems, allowing an attacker to cycle through different target identities by simply updating a software-controlled projection.

Exploiting the Digital-to-Physical Gap

To make these attacks work in the real world, researchers must account for the complex physics of how light interacts with surfaces. A 2D adversarial pattern designed on a computer will distort when projected onto a 3D object like a car or a human face.

Recent work has focused on closing this "digital-to-physical" gap. Ce Zhou and their team developed Transient Adversarial 3D Projection Attacks, which use 3D mesh modeling to optimize patterns for specific surface geometries. By accounting for ambient light and surface reflectance, they achieved near-perfect success rates in making objects "disappear" from object detectors. In the realm of biometrics, Yuanwei Liu introduced ProjAttacker, which models how light reflects off human skin to project 3D textures that fool both recognition and liveness detection.

Temporal Manipulation and Sensor Exploitation

Beyond geometry, light allows attackers to exploit the temporal mechanics of camera hardware. Most commodity cameras use a "rolling shutter," where the image is captured row-by-row rather than all at once.

Athena Sayles demonstrated that by modulating LED light at extremely high frequencies—faster than the human eye can perceive—attackers can inject adversarial striping patterns directly into the camera's frame. This Invisible Perturbations approach is particularly dangerous because the environment looks perfectly normal to a human observer, while the AI sees a corrupted, adversarial image.

Real-Time Adaptation to Motion

The most significant hurdle for physical attacks is motion. As a car drives past a sign, the angle and distance change constantly, usually breaking the adversarial effect.

To solve this, researchers are turning to closed-loop systems. Ranjie Duan developed AdvLB (Adversarial Laser Beam), which optimizes laser wavelength and geometry to maintain effectiveness against moving targets. Taking this further, the Embodied Laser Attack (ELA) uses a reinforcement learning agent to "aim" and adjust the laser in real-time based on the victim vehicle's perspective. Similarly, Amirhosein Chahe showed that mounting a screen on a vehicle to display Dynamic Adversarial Attacks allows the pattern to evolve as the distance between the attacker and victim changes, maintaining a "sweet spot" of disruption.

The Forensic Advantage

Perhaps the most unsettling aspect of light-based attacks is their lack of a physical footprint. As noted by Zhenbang Peng in a comprehensive review, projection-based attacks leave no permanent evidence on the target. Once the projector or laser is turned off, the "weapon" vanishes. This makes post-incident investigation nearly impossible, as there are no stickers to peel off or 3D-printed objects to recover. The attack exists only for the duration of the light's flight, making it a truly transient threat to the security of autonomous vision.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adversarial_attacks_using_dynamic_light__164643d7,
  title = {Adversarial Attacks Using Dynamic Light Projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-attacks-using-dynamic-light-projection/rdr_164643d7c00a4139aba158a45d8ad257},
  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_adversarial_attacks_using_dynamic_light__164643d7_v1,
  title = {Adversarial Attacks Using Dynamic Light Projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-attacks-using-dynamic-light-projection/rdr_164643d7c00a4139aba158a45d8ad257/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}