Robust Adversarial Camouflage on NonRigid Surfaces

31 concepts18 papers1 proposal

Digital adversarial perturbations are notoriously fragile. When translated into the physical world, early attempts like the printable patches developed by Simen Thys et al. demonstrated that an object detector could be blinded by a static 2D pattern. However, these early methods relied heavily on a critical, often unrealistic assumption: that the target surface remains perfectly flat and rigid.

When an adversarial pattern is printed on clothing, it encounters the chaotic physics of flexible materials. As a person walks, bends, or twists, the fabric undergoes non-rigid deformations—stretching, folding, wrinkling, and creasing. This physical variability is a long-standing hurdle in computer vision; for instance, Marco Bertini 0001 et al. noted that the non-rigid deformation of athletic apparel frequently distorts text and jersey numbers, leading to high false-positive rates in sports tracking. Even when attacks utilize everyday items like stickers, Xingxing Wei argues in Adversarial Sticker: A Stealthy Attack Method in the Physical World that highly wrinkled surfaces and extreme environmental conditions continue to pose severe challenges to real-world deployment. For adversarial attacks, the stakes are even higher: a single deep fold can disrupt the precise spatial relationships of a perturbation, instantly neutralizing its ability to fool a neural network.

Why Deformations Neutralize Adversarial Signals

To understand why wrinkles destroy adversarial attacks, we must look at how object detectors process images. Models like YOLO rely on local spatial features and texture gradients to compute "objectness" scores. A standard adversarial patch is optimized to inject a highly specific, disruptive spatial frequency into the detector's receptive field.

When fabric wrinkles, two things happen:

  1. Geometric Warping: The 2D coordinates of the pattern are compressed or stretched, shifting the adversarial pixels out of their mathematically optimized alignments.
  2. Self-Occlusion: Folds create shadows and hide sections of the pattern entirely.

As highlighted by Lifeng Huang et al. in Universal Physical Camouflage Attacks on Object Detectors, failing to account for these internal deformations means the adversarial signal loses its spatial consistency. The network simply sees a wrinkled piece of cloth rather than the carefully engineered noise designed to suppress its bounding boxes.

Bridging the Gap with Differentiable Warping

The core breakthrough in making wearable adversarial clothing lies in bringing non-rigid deformation directly into the optimization loop. Instead of optimizing a pattern for a flat surface and hoping it survives the physical world, researchers simulate the physics of fabric during training using the Expectation over Transformation (EoT) framework.

A foundational tool for this is the Thin Plate Spline (TPS) transform. Kaidi Xu et al. pioneered this approach in Adversarial T-Shirt! Evading Person Detectors in a Physical World by incorporating learned TPS mapping into their training pipeline to model how a T-shirt wrinkles during human movement. By applying random TPS warps to the adversarial texture at each optimization step, the pattern is forced to find a "universal" perturbation that remains effective across a wide distribution of geometric distortions. This approach was further validated by Zuxuan Wu et al., who combined TPS transforms with standard augmentations like scaling and rotation to bridge the digital-to-physical gap.

[2D Adversarial Texture] 
       │
       ▼
[Thin Plate Spline (TPS) Warp] ──► Simulates random wrinkles & folds
       │
       ▼
[Differentiable Renderer]       ──► Projects onto target scene
       │
       ▼
[Object Detector Loss]          ──► Backpropagates to update original texture

Similarly, Zhanhao Hu et al. utilized TPS warping within their AdvCaT framework to generate topologically plausible projections. By optimizing patterns to withstand these simulated folds, the resulting clothing maintains its deceptive properties even as the wearer moves. To capture sharper, more localized deformations, the authors of DAP: A Dynamic Adversarial Patch for Evading Person Detectors introduced a specialized "Creases Transformation" block that explicitly models fabric compression, ensuring the adversarial signal survives sharp physical folds.

From 2D Warping to 3D Physics-Based Modeling

While 2D approximations like TPS are computationally efficient, they cannot fully capture the complex 3D interactions between a moving human body and draping fabric. To achieve higher physical robustness, recent work has shifted toward full 3D physics-based simulation and neural rendering.

In AdvReal: Physical adversarial patch generation framework for security evaluation of object detection systems, Yilong Ren et al. developed a framework that uses 3D physics-based cloth modeling to simulate realistic fabric wrinkles directly onto 3D human meshes. By applying stochastic perturbations to these meshes during a joint 2D/3D optimization process, they ensure the adversarial patch remains effective across diverse postures.

Taking this a step further, Yanjie Li et al. introduced UV-Attack, which leverages dynamic Neural Radiance Fields (NeRF) combined with UV coordinate mapping. Instead of treating wrinkles as random noise, they model the continuous, non-rigid deformation of the human body and clothing as a coherent 3D space. This allows them to optimize a texture map that remains spatially consistent across complex, dynamic poses, achieving unprecedented evasion rates against state-of-the-art detectors.

Multi-Modal and Full-Body Scaling

As deformation modeling has matured, the scope of these attacks has expanded from localized patches to full-body, multi-modal camouflage. Zhanhao Hu et al. demonstrated that full-body adversarial textures (AdvTexture) naturally possess a higher resilience to local folds because the adversarial signal is distributed across the entire garment, preventing localized wrinkles from neutralizing the attack.

Furthermore, this paradigm has been extended beyond visible light. Xiaopei Zhu et al. combined digital fabric deformation simulations with specialized aerogel insulation to create infrared invisible clothing, proving that non-rigid robustness translates to thermal surveillance systems. Most recently, Guixu Lin et al. extended these principles to event-based pedestrian detectors, optimizing high-contrast textures that exploit the motion-sensitive nature of neuromorphic cameras despite the complex geometry of moving garments.

Key Foundational Work

Separating Reflection and Transmission Images in the Wild (2018) — contributed 1 concept elements to this direction.
Automatic Annotation of Sport Video Content (2005) — 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_non_rigid_surface_deformation_a805a071,
  title = {non-rigid surface deformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-rigid-surface-deformation/rdr_a805a071037547d29e8c9d7f0073b5b1},
  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_non_rigid_surface_deformation_a805a071_v1,
  title = {non-rigid surface deformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-rigid-surface-deformation/rdr_a805a071037547d29e8c9d7f0073b5b1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake