Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Signal Processing

Background:

• 3D point cloud learning models are crucial for safety-critical applications.
• These models are vulnerable to adversarial attacks, which can compromise their reliability.
• Existing attack methods often perturb point clouds directly in the data space, potentially ignoring geometric properties.

Purpose of the Study:

• To propose a novel point cloud attack method operating in the graph spectral domain.
• To investigate perturbing graph transform coefficients to manipulate geometric structures.
• To enhance adversarial attack effectiveness and imperceptibility in 3D point cloud models.

Main Methods:

• Adaptive transformation of point coordinates to the spectral domain using Graph Fourier Transform (GFT).
• Analysis of spectral band influence on geometric structure.
• Perturbation of GFT coefficients via a learnable graph spectral filter with a low-frequency constraint.
• Generation of adversarial point clouds through inverse GFT.

Main Results:

• The proposed graph spectral domain attack effectively perturbs geometric structures.
• The attack achieves high success rates in generating adversarial examples.
• Perturbations are constrained to imperceptible high-frequency components, maintaining visual fidelity.

Conclusions:

• The graph spectral domain attack offers a new perspective for generating adversarial point clouds.
• This method outperforms traditional point-wise perturbation by considering geometric characteristics.
• The approach demonstrates significant potential for evaluating the robustness of 3D point cloud models.