Core Concepts
The author aims to fill the gap in transferable targeted 3D adversarial attacks by proposing a novel framework, TT3D, that enhances black-box transferability and naturalness through dual optimization in the grid-based NeRF space.
Abstract
The content discusses the development of transferable targeted 3D adversarial attacks, highlighting the significance of such attacks in security-critical tasks. The proposed framework, TT3D, utilizes dual optimization in the grid-based NeRF space to enhance transferability and naturalness. Experimental results demonstrate superior cross-model transferability and adaptability across different renders and vision tasks. Additionally, 3D adversarial examples are produced using 3D printing techniques for real-world validation under various scenarios.
Key points include:
Importance of transferable targeted 3D adversarial attacks for security-critical tasks.
Introduction of TT3D framework for generating transferable targeted 3D adversarial examples.
Dual optimization strategy targeting both feature grid and MLP parameters in the grid-based NeRF space.
Experimental results showing superior cross-model transferability and adaptability.
Production of 3D adversarial examples with 3D printing techniques for real-world validation.
Stats
ResNet-101: Attack success rate - 88.98%
DenseNet-121: Attack success rate - 96.74%
Quotes
"Crafting such transferable targeted attacks is particularly challenging because they not only need to achieve a specific misclassification but also must avoid overfitting."
"We propose a novel framework called TT3D for generating transferable targeted 3D adversarial examples."
"Our method also exhibits visual naturalness compared to mesh-based optimization methods."