Transferability of Fast Gradient Sign Method

Abstract

Image classification in computer vision is a process which classifies an image depending on its content. While classifying an object is trivial for humans, robust image classification is still a challenge in computer vision applications. The robustness of such image classification models in real world applications is a major concern. Adversarial examples are specialized inputs created with the purpose of confusing a classifier, resulting in the misclassification of a given input. Some of these adversarial examples are indistinguishable to humans, but the classifier can still be tricked into outputting the wrong class. In some cases adversarial examples can be transferred: an adversarial example crafted on a target model fools another model too. In this paper we evaluate the transferability of adversarial examples crafted with Fast Gradient Sign Method across models available in the open source Tensorflow machine learning platform (using ResNetV2, DenseNet, MobileNetV2 and InceptionV3).