Box Attack

Box attacks, encompassing both white-box and black-box variations, and the more challenging "no-box" attacks, investigate the vulnerability of deep neural networks (DNNs) to adversarial perturbations. Current research focuses on developing efficient attack algorithms, such as variations of projected gradient descent (PGD), and exploring novel defense mechanisms like vector quantization. These studies are crucial for understanding and mitigating the risks posed by adversarial examples to the reliability and security of DNNs across diverse applications, including image classification, human action recognition, and 3D point cloud analysis. The development of effective no-box attacks, requiring no knowledge of the target model, highlights the need for robust and generalizable defenses.

Papers

May 8, 2024

Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations
Xuyang Zhong, Yixiao Huang, Chen Liu
Global Evaluation Adversarial Perturbation Efficient Training Robust Model Sparse Attack Sparse Perturbation Box Attack

August 10, 2023

Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient
Zhengzhi Lu, He Wang, Ziyi Chang, Guoan Yang, Hubert P. H. Shum
Adversarial Attack Gradient Based Attack Skeleton Based Human Activity Recognition Box Attack

May 23, 2023

Adversarial Defenses via Vector Quantization
Zhiyi Dong, Yongyi Mao
Vector Quantization Adversarial DEfense Box Attack

October 19, 2022

No-Box Attacks on 3D Point Cloud Classification
Hanieh Naderi, Chinthaka Dinesh, Ivan V. Bajic, Shohreh Kasaei
Adversarial Attack Point Cloud Classification PointNet Model Box Attack

March 9, 2022

Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation
Qilong Zhang, Chaoning Zhang, Chaoqun Li, Jingkuan Song, Lianli Gao
Adversarial Perturbation Black Box Adversarial Vulnerability Box Attack

Box Attack

Papers

Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations

Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient

Adversarial Defenses via Vector Quantization

No-Box Attacks on 3D Point Cloud Classification

Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation