Existing GenAttack is a state-of-the-art methods on fooling deep learning based image classifier in black-box setting. However, it still requires probability after softmax layer for adversarial samples selection. In this work, we extend the GenAttack to hard-label case. With an appropriate amount of queries to the TOP-1 class label, the GenAttack can attack the Inception based image classifier on ImageNet.
L2 distance changes w.r.t. the number of queries
Fooling the image classifier by classifying squirrel as parking meter