On the Generalization of GAN Image Forensics

Abstract

Recently GAN generated face images are more and more realistic with high-quality, even hard for human eyes to detect. On the other hand, the forensics community keeps on developing methods to detect these generated fake images and try to ensure the credibility of visual contents. Although researchers have developed some methods to detect generated images, few of them explore the important problem of generalization ability of forensics model. As new types of GANs are emerging fast, the generalization ability of forensics models to detect new types of GAN images is absolutely an essential research topic, which is also very challenging. In this paper, we explore this problem and propose to use preprocessed images to train a forensic CNN model. By applying similar image level preprocessing to both real and fake images, unstable low level noise cues are destroyed, and the forensics model is forced to learn more intrinsic features to classify the generated and real face images. Our experimental results also prove the effectiveness of the proposed method.

Incremental learning for the detection and classification of GAN-generated images

“…The first one is XceptionNet, the same architecture used in our iCaRL experiments, that shows good results in [14] for GAN detection. Then, we compared the two training procedures proposed in [16], where the discriminator of DCGAN [24] is used as classifier and two different pre-processing procedures are used during the training phase, that is, Gaussian noise and Gaussian blur. We will refer to them as M Gn and M Gb , respectively.…”

“…Approaches based on convolutional neural networks have proven to be very effective. Several architectures have been proposed so far [14], [15], [16] showing a very good accuracy in detecting GAN-generated images, even after compression. The main problem is that new GAN architectures for generating synthetic data are proposed by the day, requiring the detector to be either re-trained on larger and larger training sets, or fine-tuned on them.…”

A Comprehensive Review of Deep-Learning-Based Methods for Image Forensics

“…Meanwhile in [ 158 ], authors exploited the inconsistencies between warped face area and the surrounding background. The method in [ 159 ] adopted Gaussian noise extraction as a pre-processing step for a CNN, enforcing the network to learn more meaningful features about GAN traces.…”

Deep visual identity forgery and detection

“…中山大学骆伟祺教授团队 [83] 提出利用高通 滤波器进行预处理以提高检测效果. 中国科学院董晶教授团队 [84] 指出可以借助高斯 (Gauss) 模糊和 增加高斯噪声等图像预处理操作来减少伪造图像和真实图像中存在的不稳定高频失真, 从而有助于学 习更加关键的伪造特征以提高伪造检测的性能.…”