Reproduce “Less Is More: Sparse and Cooperative Perturbation for Point Cloud Attacks” on OpenResearch

About this paper

Keke Tang, Tianyu Hao, Xiaofei Wang, Weilong Peng, Denghui Zhang, Peican Zhu, Zhihong Tian

Researchers from Guangzhou University and collaborators developed Sparse and Cooperative Perturbation (SCP), an adversarial attack framework that identifies and leverages cooperative interactions among a minimal subset of points in 3D point clouds. This method achieves a 100% attack success rate on various state-of-the-art 3D classifiers by perturbing two orders of magnitude fewer points than dense attacks, resulting in superior imperceptibility and physical realizability.

Key insights

• Point cloud DNNs exhibit strong nonlinear interactions, meaning jointly perturbing specific points can yield a synergistic adversarial effect greater than individual perturbations.
• "Cooperative behavior" of perturbed points is linked to the local strict convexity of the misclassification loss function, mathematically identifiable through the positive-definiteness of the Hessian matrix.
• Identifying and leveraging these cooperative subsets enables the creation of highly effective adversarial examples with unprecedented sparsity and imperceptibility, posing a deeper threat than individual point attacks.

Summary by alphaXiv.