A comparative study of white-box and black-box adversarial attacks to the deep neural networks with different architectures

Abstract

A few years ago, it was discovered that the Deep Convolutional Neural Networks (CNN) are vulnerable to so-called adversarial attacks. An adversarial attack supposes a subtle modification of an original image in such a way that the changes are almost invisible to the human eye. In this work, we are concentrating on biomedical images, which are playing the key role in the disease diagnosis and monitoring of various treatment processes. We present detailed results on the success rate for both white-box and black-box untargeted attacks to five types of popular deep CNN architectures including InceptionV3, Xception, ResNet50, DenseNet121, and Mobilenet