LPF-Defense: 3D adversarial defense based on frequency analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• 3D point clouds are vital in safety-critical applications.
• Deep neural networks (DNNs) process 3D point clouds but are vulnerable to adversarial attacks.
• Attacks exploit DNN overreliance on irrelevant features in training data.

Purpose of the Study:

• To propose a defense framework against 3D adversarial attacks on deep learning models for point clouds.
• To improve the robustness of DNNs by removing unnecessary information from training data.

Main Methods:

• Introduced the LPF-Defense framework to suppress high-frequency content in training data.
• Analyzed adversarial perturbations within the high-frequency components of adversarial point clouds.
• Evaluated defense performance against six adversarial attacks on PointNet, PointNet++, and DGCNN.

Main Results:

• LPF-Defense achieves state-of-the-art defense performance against multiple adversarial attacks.
• Demonstrated significant accuracy improvements on synthetic (ModelNet40, ShapeNet) and real (ScanObjectNN) datasets.
• Achieved average accuracy increases of 3.8% (Drop100) and 4.26% (Drop200) over existing methods.

Conclusions:

• Suppression of high-frequency content is an effective defense against 3D adversarial attacks.
• LPF-Defense enhances model robustness and classification accuracy on original datasets.
• An open-source implementation is available to facilitate further research.