Detection of adversarial examples based on similarity between object external features described in natural language and images

Abstract

Detection of adversarial examples is an important issue. This research focuses on the fact that the adversarial example looks similar to the original image and the appearance features of the object. Through experiments on Cifar-10 and Tiny-ImageNet, we have shown that the proposed method achieves a certain detection performance in the black box setting, and we confirmed that Tiny-ImageNet performs better when using object appearance features to detect hostile samples.We also confirmed that the performance of the no-box setting is not so different from that of the black-box setting.