Enhancing Adversarial Examples with Flip-Invariance and Brightness-Invariance

Abstract

Despite of achieving remarkable success in computer vision tasks, convolutional neural networks still face the threat of adversarial examples, crafted by adding small human-invisible perturbations on clean inputs. Usually, most of existing black-box adversarial attacks show extremely low transferability while encountering powerful defense models. In this paper, based on the observed invariant property of convolutional neural networks (i.e., the models could maintain accuracy to transformed images), we propose two new methods to improve the transferability which are called as the flip-invariant attack method (FIM) and the brightness-invariant attack method (BIM), respectively. Both the novel approaches derive multiple different logit outputs by inputting the transformed copies of the original image into the white-box model. Simultaneously, the ensemble of these outputs is attacked to avoid overfitting the white-box model and generating more transferable adversarial examples. Moreover, the newly-proposed FIM and BIM methods can be naturally combined with other gradient-based methods. Extensive experiments on the ImageNet dataset prove that our methods achieve higher attack success rate and higher transferability than previous gradient-based attack methods.