A Policy-Driven Black-Box Adversarial Example With Location Optimization Against 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning Security

Background:

• Adversarial attacks are crucial for 3D object detection security.
• White-box attacks require full model access, limiting large-scale applications.

Purpose of the Study:

• To propose a novel Policy-Driven Black-box Attack (BAT) for 3D object detection.
• To optimize attack locations without detailed knowledge of victim models.

Main Methods:

• Introduced a density-aware pattern generator for scene-adaptive attack clusters.
• Utilized deep deterministic policy gradient (DDPG) in deep reinforcement learning to train an attack agent.
• Employed joint critic and actor loss for iterative optimization of attack locations.

Main Results:

• BAT effectively reduces the accuracy of prominent 3D object detection models.
• Achieved significantly enhanced attack success rates, surpassing state-of-the-art methods.
• Increased attack transferability by 20 times compared to existing black-box approaches.

Conclusions:

• BAT establishes a new baseline for reinforcement learning-based black-box adversarial attacks in 3D object detection.
• Demonstrates the efficacy of DDPG for optimizing adversarial attacks in point cloud applications.
• Highlights the potential for robust security evaluations in autonomous driving systems.