Abstract
Nowadays, 3D data plays an indelible role in the computer vision field. However, extensive studies have proved that deep neural networks (DNNs) fed with 3D data, such as point clouds, are susceptible to adversarial examples, which aim to misguide DNNs and might bring immeasurable losses. Currently, 3D adversarial point clouds are chiefly generated in three fashions, i.e., point shifting, point adding, and point dropping. These point manipulations would modify geometrical properties and local correlations of benign point clouds more or less. Motivated by this basic fact, we propose to defend such adversarial examples with the aid of 3D steganalysis techniques. Specifically, we first introduce an adversarial attack and defense model adapted from the celebrated Prisoners' Problem in steganography to help us comprehend 3D adversarial attack and defense more generally. Then we rethink two significant but vague concepts in the field of adversarial example, namely, active defense and passive defense, from the perspective of steganalysis. Most importantly, we design a 3D adversarial point cloud detector through the lens of 3D steganalysis. Our detector is double-blind, that is to say, it does not rely on the exact knowledge of the adversarial attack means and victim models. To enable the detector to effectively detect malicious point clouds, we craft a 64-D discriminant feature set, including features related to first-order and second-order local descriptions of point clouds. To our knowledge, this work is the first to apply 3D steganalysis to 3D adversarial example defense. Extensive experimental results demonstrate that the proposed 3D adversarial point cloud detector can achieve good detection performance on multiple types of 3D adversarial point clouds.
Abstract (translated)
URL
https://arxiv.org/abs/2205.08738
