Geometry Preserving Adversarial Perturbations for 3D Point Clouds

This direction is merged into another direction: surface-constrained perturbation. Its last version stays readable and citable here.

Version 1 of 125 papers29 conceptsfrozen 2026-10-04

In 3D computer vision, point clouds are the primary language for representing the physical world. From the LiDAR sensors on autonomous vehicles to the depth cameras in Augmented Reality (AR) headsets, deep learning models rely on these collections of $(x, y, z)$ coordinates to recognize objects. However, these models are vulnerable to adversarial attacks—subtle, malicious modifications that trigger misclassification.

Early 3D attacks were "point-wise": they treated every point as an independent variable. While mathematically effective, these attacks produced "jittery" surfaces and floating outliers that were easily spotted by humans and neutralized by simple statistical filters. The research direction has since shifted toward geometry-preserving perturbations, where the goal is to hide the adversarial signal within the natural structural flow of the object.

The Failure of Point-wise Noise

Traditional adversarial attacks on 3D data often rely on shifting individual points in directions that maximize model error. As Kibok Lee 0003, Zhuoyuan Chen, and colleagues noted in ShapeAdv, this approach creates high-frequency noise. Because real-world sensors rarely produce such jagged artifacts, these attacks are brittle; a simple statistical outlier removal (SOR) filter can "clean" the point cloud before it ever reaches the neural network.

Furthermore, humans are remarkably sensitive to changes in point density. Even if the displacement of points is small, local "clumping" or "thinning" creates a visual signature of tampering. As Keke Tang and Lujie Huang argue in FLAT, preserving regional uniformity is often more important for imperceptibility than minimizing the total distance points are moved.

Attacking the Manifold

To bypass both human eyes and geometric filters, researchers began treating point clouds not as discrete dots, but as samples of a continuous 2D surface (a manifold) embedded in 3D space.

In Deep Manifold Attack, Keke Tang and Weilong Peng proposed "stretching" the underlying parameter plane of a generative model. Instead of moving points, they deform the surface itself. This ensures the resulting adversarial object remains smooth and realistic. Similarly, Manifold Constraints for Imperceptible Adversarial Attacks uses invertible auto-encoders to ensure that perturbations stay "on-manifold," making them indistinguishable from the original geometric structure.

Frequency and Spectral Attacks

Another way to preserve geometry is to move the attack into the Graph Spectral Domain. Just as an image can be decomposed into low and high frequencies, a 3D shape has a "spectral" representation based on its graph Laplacian.

Daizong Liu and Wei Hu introduced GSDA, which perturbs the Graph Fourier Transform (GFT) of the point cloud. By focusing the attack on specific frequency components and enforcing low-frequency constraints, they create deformations that look like natural shape variations rather than sensor noise. This approach is particularly effective because many defenses, such as those described in GSDA++, attempt to defend models by training them to ignore high-frequency noise—a defense that spectral attacks naturally circumvent.

Structural Deformations: Cages and Skeletons

The most recent frontier in this area borrows techniques from computer graphics to manipulate the "skeleton" or "envelope" of a shape.

  1. Medial Axis Transform (MAT): In MAT-Adv, Keke Tang and colleagues represent a shape as a collection of spheres. By perturbing this "skeleton," the attack induces global, structurally coherent changes that are highly transferable across different AI models.
  2. Cage-Based Deformation: Cage-Based Deformation for Transferable and Undefendable Point Cloud Attack places the point cloud inside a coarse "cage." Moving the vertices of the cage smoothly deforms the thousands of points inside it. Because the deformation is mathematically smooth, it bypasses almost all geometric denoising filters.

Why It Works: The Intuition of Realism

The core intuition behind this research direction is that adversarial signals are most effective when they mimic legitimate data variance.

When an attack like Isometric 3D Adversarial Examples by Yibo Miao and Yinpeng Dong preserves Gaussian curvature, it is essentially performing a "non-rigid" transformation—similar to how a person might bend their arm or how a plastic bottle might slightly dent. Neural networks, which are trained to be invariant to some of these changes but sensitive to others, can be easily led astray by a perturbation that looks like a natural, physical transformation but is actually a calculated strike against the model's decision boundaries.

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_geometry_preserving_adversarial_perturba_905d0b4b,
  title = {Geometry Preserving Adversarial Perturbations for 3D Point Clouds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometry-preserving-adversarial-perturbations-for-3d-point-clouds/rdr_905d0b4b4d1442de85a20f2dd2a29bf4},
  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_geometry_preserving_adversarial_perturba_905d0b4b_v1,
  title = {Geometry Preserving Adversarial Perturbations for 3D Point Clouds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometry-preserving-adversarial-perturbations-for-3d-point-clouds/rdr_905d0b4b4d1442de85a20f2dd2a29bf4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}