Generating Adversarial Examples in Limited Queries with Image Encoding and Noise Decoding

摘要

Deep neural network (DNN) has been widely used in many application scenarios, but it is easy to be affected by slight disturbance, which is referred to as adversarial examples, and produces wrong decisions endangering system security. It is difficult to obtain the structure and parameter information of the target network in the actual scene. In this paper, we propose an optimization-based method to generate adversarial examples. Inspired by the autoencoder and Conditional Generative Adversarial Net (CGAN), we train the encoder to encode the image into latent vector, of which the specific region (category field) is used to represent the category information of the image, and then train the decoder to generate adversarial examples of the specified category. We perform Natural Evolution Strategy in the low-dimensional latent space to further improve the attack. The result shows that our method can successfully attack the target network within only one or a few queries. We evaluate our method on Sign-Language-Digits (SLD) and Cifar-10 datasets. Compared with other methods, our approach can attack the target model with less queries, and maintain a high success rate.