Geometry-Aware Generation of Adversarial Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Machine learning models are vulnerable to adversarial examples, with most research focused on 2D images.
• Existing 3D point cloud adversarial methods often produce noticeable outliers, making them easy to defend against.
• Human perception of 2D images and 3D shapes differs, suggesting unique approaches for 3D adversarial attacks.

Purpose of the Study:

• To develop a novel method for generating more effective and imperceptible adversarial examples for 3D point cloud data.
• To address the limitations of existing 3D adversarial attack methods, specifically the issue of point outliers.
• To leverage geometric properties of 3D shapes for improved adversarial attack generation.

Main Methods:

• Proposed Geometry-Aware Adversarial Attack (GeoA3) using geometry-aware objectives focusing on surface smoothness and fairness.
• Employed a targeted attack misclassification loss for generating increasingly malicious adversarial signals.
• Introduced Geo+A3-IterNormPro with Iterative Normal Projection (IterNorPro) for surface-level adversarial attacks.

Main Results:

• GeoA3 generates adversarial point clouds that are more harmful and harder to defend against.
• Adversarial examples produced by GeoA3 are imperceptible to humans, unlike those from previous methods.
• Quantitative and qualitative evaluations demonstrated the superiority of the proposed methods over existing techniques.

Conclusions:

• Geometry-aware objectives are effective in generating high-quality, imperceptible adversarial examples for 3D point clouds.
• The proposed GeoA3 method represents a significant advancement in 3D adversarial attack research.
• The developed techniques offer improved robustness against outlier-based defenses and enhance the stealthiness of adversarial attacks.