Toward leveraging intrinsic point cloud features in 3D adversarial attacks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) face adversarial attacks, especially with 3D point cloud data.
• Identifying critical points is key for generating effective adversarial attacks.

Purpose of the Study:

• To understand which 3D point cloud features predict adversarial points.
• To develop a novel adversarial attack method based on intrinsic point cloud characteristics.

Main Methods:

• Defined fourteen key point cloud features (e.g., edge intensity, distance from centroid).
• Employed random forest and multiple linear regression to assess feature predictive power.
• Designed and tested a new attack method across four DNN architectures (PointNet, PointNet++, DGCNN, PointConv).

Main Results:

• The proposed attack method shows generalizability and improved transferability across different DNN architectures.
• Achieved higher success rates (2-4%) in transfer settings compared to random guessing.
• Demonstrated that intrinsic point cloud features can be leveraged for adversarial attacks.

Conclusions:

• Shifts focus from model-specific attacks to data-driven, feature-based methods.
• Offers potential for reduced computational costs and more interpretable deep learning models.
• Provides insights for model explainability, robust learning, and defense mechanisms against adversarial threats.