Imperceptible Transfer Attack and Defense on 3D Point Cloud Classification.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Limited research exists on 3D model vulnerabilities compared to 2D images.
• Existing 3D attacks often create perceptible distortions and suffer from poor transferability to black-box models.

Purpose of the Study:

• To propose a novel Imperceptible Transfer Attack (ITA) for 3D point clouds.
• To enhance the imperceptibility and transferability of adversarial attacks on 3D models.
• To develop a robust defense strategy for 3D models against ITA.

Main Methods:

• Constraining point perturbation along normal vectors for imperceptibility.
• Developing an adversarial transformation model to improve attack transferability.
• Training robust black-box 3D models using discriminative point cloud representations for defense.

Main Results:

• The proposed ITA demonstrates superior imperceptibility and transferability over existing methods.
• Extensive evaluations confirm the effectiveness of the ITA.
• The defense strategy significantly enhances 3D model robustness against sophisticated attacks.

Conclusions:

• The ITA presents a significant advancement in adversarial attacks on 3D point clouds.
• The developed defense mechanism offers a promising solution for securing 3D models.
• This work highlights the critical need for robust 3D model security.