Random Transformation of image brightness for adversarial attack

Deep neural networks (DNNs) are vulnerable to adversarial examples, which crafted by adding small, human-imperceptible perturbations the original images, but make model output inaccurate predictions. Before DNNs deployed, attacks can thus be an important method evaluate and select robust models in safety-critical applications. However, under challenging black-box setting, attack success rate, i.e., transferability of still needs improved. Based on image augmentation methods, this paper found that random transformation brightness eliminate overfitting generation examples improve their transferability. In light phenomenon, proposes example method, integrated with Fast Gradient Sign Method (FGSM)-related methods build a more gradient-based generate better Extensive experiments ImageNet dataset have demonstrated effectiveness aforementioned method. Whether normally or adversarially trained networks, our has higher rate for than other based data augmentation. It is hoped help robustness models.