Dual-targeted adversarial noise for 3D point cloud classification model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are increasingly used for 3D point cloud recognition.
• These models are vulnerable to adversarial attacks that exploit data structures.
• Existing adversarial methods may not effectively target multiple distinct classes across different models.

Purpose of the Study:

• To introduce a novel method for generating dual-target adversarial examples in 3D point cloud data.
• To cause different deep learning models to misclassify into distinct, attacker-specified classes.
• To enhance the security and robustness of 3D point cloud recognition systems.

Main Methods:

• Developed a method to generate dual-target adversarial examples for point cloud data.
• Utilized feedback from multiple models to minimize the loss function.
• Ensured targeted misclassification into distinct attacker-specified classes.
• Validated the approach on the ModelNet40 dataset using PointNet and PointNet++ models.

Main Results:

• Achieved high attack success rates: 99.8% for PointNet and 84.16% for PointNet++.
• Demonstrated effectiveness through visualization of attack success rates, distortions, and point clouds.
• Successfully generated dual-target adversarial examples causing distinct misclassifications.

Conclusions:

• The proposed method is effective in generating dual-target adversarial examples for 3D point cloud recognition.
• The technique has significant implications for security, especially in adversarial contexts like military applications.
• Further research can explore more complex scenarios and model architectures.