Physically Realizable Adversarial Creating Attack against Vision-based BEV Space 3D Object Detection.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Cybersecurity

Background:

• Vision-based 3D object detection is vital for autonomous driving but vulnerable to adversarial attacks.
• Existing attacks on 2D detectors are less effective against 3D detectors due to viewpoint and rendering challenges.

Purpose of the Study:

• To propose a novel adversarial attack targeting vision-based 3D object detectors.
• To demonstrate the effectiveness of a single, universal 2D adversarial poster placed on the road surface.

Main Methods:

• Developed an image-3D algorithm for differentiable rendering of 3D space onto 2D images.
• Employed a ground-truth masked optimization strategy to train the adversarial poster.
• Validated the attack through extensive simulations and real-world experiments.

Main Results:

• A single 2D adversarial poster can successfully deceive 3D object detectors.
• The attack is universal, effective across diverse scenes, viewpoints, and distances.
• Real-world experiments confirmed the attack's efficacy.

Conclusions:

• The proposed adversarial poster attack presents a significant threat to vision-based 3D object detection in autonomous driving.
• Further research into transferability and defense strategies is warranted to mitigate this vulnerability.