Abstract:

Since the birth of generative adversarial networks (GANs), the research on it has become a hot spot in the field of machine learning. It uses the mechanism of adversarial learning to train model solving the problem that the generation algorithm cannot solve. Due to the advantages of GANs, researchers conduct in-depth research on it and  a large number of derivative models of GANs are produced, which empowers the rapid development of GANs and the formation of so-called GAN-Zoo. GANs is widely used in visual field, audio field, natural language field and other fields, such as image generation, image translation, text generation, audio conversion, natural language translation and so on. Based on the traditional GANs, this paper introduces and summarizes the prominent aspects of GANs research in recent years. Firstly, this paper introduces the basic theory of GANs,  summarizes the main derivative models of GANs in recent years, and finally summarizes the main application results of GANs in image field and information security field.

Key words: generative adversarial networks (GANs), divergence function, neural networks, generative model

摘要：

自生成对抗网络（GANs）诞生以来，对其研究已经成为机器学习领域的一个热点。它利用对抗学习的机制训练模型，解决了当年生成算法无法解决的问题。由于GANs的优势，研究者们对其进行深入的研究，产生了许多GANs的衍生模型，这使得GANs得到了快速的发展，形成了所谓的GAN-Zoo。GANs被广泛应用于视觉领域、音频领域、自然语言领域及其他各种领域中，如图像生成、图像翻译、文本生成、音频转换和自然语言翻译等。从传统GANs出发，对近几年内GANs的研究中较为突出的方面进行总结，首先介绍了传统GANs的基本理论，然后对近年来GANs的主要衍生模型进行分析，最后总结了GANs在图像领域和信息安全领域中的主要应用成果。

关键词: 生成对抗网络（GANs）, 散度函数, 神经网络, 生成模型