Exploring Adversarial Robustness of LiDAR Semantic Segmentation in Autonomous Driving

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Deep learning excels in 2D/3D vision but is vulnerable to adversarial attacks.
• Adversarial attacks on autonomous vehicle perception are increasingly studied, yet 3D LiDAR semantic segmentation remains underexplored.

Purpose of the Study:

• To investigate the adversarial robustness of 3D LiDAR semantic segmentation in autonomous vehicles.
• To develop and analyze LiDAR point-based adversarial attack methods.

Main Methods:

• Developed and analyzed three LiDAR point-based adversarial attack methods.
• Evaluated attacks on various networks using the SemanticKITTI dataset.
• Investigated class-wise point distribution's influence on adversarial robustness.

Main Results:

• The Cylinder3D network exhibited the highest susceptibility to the analyzed adversarial attacks.
• Ground-level points were found to be particularly vulnerable to point perturbation attacks.
• Networks utilizing point data representations showed notable resistance to attack transferability.

Conclusions:

• Adversarial attacks pose a significant threat to 3D LiDAR semantic segmentation in autonomous vehicles.
• Understanding class-wise vulnerability, especially for ground-level points, is crucial for developing robust systems.
• Findings provide a foundation for creating advanced adversarial attacks and effective countermeasures for LiDAR-based perception.