引用本文
  • 侯贵洋,赵险峰,曹 纭,何晓磊.基于复杂纹理区域伪造痕迹的GAN生成图像检测方法[J].信息安全学报,已采用    [点击复制]
  • HOU Guiyang,ZHAO Xianfeng,CAO Yun,HE Xiaolei.GAN-Generated image detection method based on forged traces in complex texture region[J].Journal of Cyber Security,Accept   [点击复制]
【打印本页】 【在线阅读全文】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

过刊浏览    高级检索

本文已被:浏览 335次   下载 0  
基于复杂纹理区域伪造痕迹的GAN生成图像检测方法
侯贵洋, 赵险峰, 曹 纭, 何晓磊
0
(中国科学院信息工程研究所 信息安全国家重点实验室 北京中国)
摘要:
近年来,随着生成对抗网络(Generative adversarial network,GAN)技术的快速发展,GAN生成图像的真实感得到了极大的提高。由于人脸图像和场景图像在金融、社交、文化以及政治等领域发挥着极其重要的作用,GAN生成的人脸图像和场景图像可能会给社会带来巨大的潜在威胁。因此,急需有效的方法来对GAN生成图像进行检测。现有的生成对抗网络在生成图像时,在图像的各个位置都有可能留下伪造痕迹,通过对生成模型潜在空间以及生成图像拉普拉斯金字塔进行分析发现,生成模型更可能在图像的复杂纹理区域留下伪造痕迹,即生成模型很难完美地重建图像中复杂的纹理。现有的检测方法普遍基于卷积神经网络(Convolutional neural network,CNN)构建,然而CNN存在倾向于从图像的有限区域检查伪造、感受野不够灵活以及对伪造的理解与人类不同(人类通常会在整张图像发现代表性伪造)的缺点。因此本文提出了一种新颖的GAN生成图像检测框架,即在纹理注意力机制和通道注意力机制的指导下,采用Transformer编码模型融合图像局部特征和全局特征进行分类检测。在人脸数据集和场景数据集上的实验结果表明,本文提出的方法在多种图像生成类型(人脸图像生成、翻译、风格迁移)上的检测准确率均优于当前最先进的检测方法。最后,提出了将模型思想推广到一般任务上的方法,论述了模型思想具有较大的推广价值,并展望了未来可能的研究方向。
关键词:  纹理注意力机制  通道注意力机制  特征融合  Transformer编码  多种生成类型检测  模型推广
DOI:
投稿时间:2021-10-01修订日期:2021-12-16
基金项目:国家自然科学基金项目(面上项目,重点项目,重大项目)、国家重点研发计划
GAN-Generated image detection method based on forged traces in complex texture region
HOU Guiyang, ZHAO Xianfeng, CAO Yun, HE Xiaolei
(State Key Laboratory of Information Security,Institute of Information Engineering,Chinese Academy of Sciences)
Abstract:
In recent years, with the rapid development of Generative Adversarial Networks (GAN) technology, the realism of GAN-Generated images has been greatly improved. Because face images and scene images play an extremely important role in financial, social, culture, political and other fields, the face images and scene images generated by GAN may pose a huge potential threat to society. Therefore, there is an urgent need for effective methods to detect GAN-Generated images. When the existing generative adversarial network generates whole images, it is possible to leave traces of forgery in arbitrary position of the generated images. Through analyze the potential space of generation model and Laplacian pyramid of generated images, it is found that the generation model is more likely to leave traces of forgery in the complex texture area, that is, it is difficult for generation model to perfectly reconstruct the complex texture in the image. Existing detection methods are generally based on Convolutional Neural Network(CNN). However, CNN tend to check forgery from a limited area of the image, receptive field is not flexible enough and understanding of forgery is different from that of humans(humans usually find representative forgery over the entire image). Therefore, this paper proposed a novel GAN-Generated image detection framework, that is, under the guidance of the texture attention mechanism and the channel attention mechanism, Transformer coding model fuse local and global features of image for classification detection. The experimental results on the face dataset and the scene dataset show that the detection accuracy of the proposed method is superior to the current state-of-the-art method in a variety of image generation types (face image generation, translation, style transfer). Finally, a method to extend the model idea to general tasks is proposed, discusses that the model idea has great promotion value, and looks forward to the possible future research directions.
Key words:  texture attention mechanism  channel attention mechanism  feature fusion  Transformer encoder  multiple generation type detection  model promotion