Generating Adversarial Examples Based on Subarea Noise Texture for Efficient Black-Box Attacks

Abstract

Nowadays, machine learning algorithms play a vital role in the field of artificial intelligence. However, it has been proved that deep convolutional networks (DCNs) are vulnerable to interference from adversarial examples. In this paper, we innovatively simulate natural textures by adding continuous noise to image subareas to generate adversarial examples, which can achieve up to 90% fooling rate on the object detection tasks (YOLOv3/Inceptionv3). The experimental results show that DCNs based on ImageNet dataset training relies too much on the feature aggregation of lower subareas in the classification task. It is instructive that when training DCNs, we need to consider not only the pursuit of accuracy but also the nature of model feature learning.