Research on Generating Adversarial Examples in Applications

The deep neural network is a highly expressive model, which plays an extremely critical role in modern artificial intelligence applications. Adversarial samples can change the prediction results of neural networks by imposing imperceptible perturbations on the original images, which brings new challenges to deep learning. In this paper we summarize the methods of generating adversarial samples in recent years, intuitively feel the development of adversarial samples from the time of publication, and briefly classify them from the perspective of algorithm principles. At the same time, for practical applications, the advantages and disadvantages of the algorithm are analyzed and summarized, and the conditions for the algorithm to be suitable for application are proposed: no need to know the specific structure of the algorithm, high quality of the adversarial samples and convenient migration. The analysis points out that among the typical methods, MI-FGSM, ONE-PIXEL, and methods using GAN are more practical and worthy of further study.