面向深度伪造的溯源取证方法

Analyzing deepfake provenance and forensics

近年来,生成对抗网络(GAN)的迅速发展使得合成图像越来越逼真,对个人和社会造成了极大的威胁。现有的研究致力于被动地鉴别伪造产品,但在真实应用场景下通常面临通用性不足和鲁棒性差等两大难题。因此,该文提出了一种面向深度伪造的溯源取证方法,将秘密信息隐藏到图像中以追踪伪造图像的源头。设计了一个端到端的深度神经网络,该网络由嵌入网络、 GAN模拟器和恢复网络等3部分组成。其中,嵌入网络和恢复网络分别用于实现秘密信息的嵌入和提取,GAN模拟器用于模拟各种GAN的图像变换。实验中在已知GAN的篡改下恢复图像的平均归一化互相关(NCC)系数高于0.9,在未知GAN的篡改下平均NCC也能达到0.8左右,具有很好的鲁棒性和通用性。此外,该方法中嵌入的秘密信息具有较好的隐蔽性,平均峰值信噪比(PSNR)在30 dB左右。

In recent years, the rapid development of generative adversarial networks(GAN) has made synthesized images more and more realistic, which poses great threats to individuals and society. Existing research has focused on passively identifying deepfakes, but real-world applications are usually insufficiently general and robust. This paper presents a method for deepfake provenance and forensics. Deepfakes hide secret information in facial images to track the source of the forged image. An end-to-end deep neural network was designed to include an embedding network, a GAN simulator, and a recovery network. The embedding network embeds the secret information in the picture while the recovery network extracts the information. The GAN simulator simulates various GAN-based image transformations. The average normalized cross correlation coefficient(NCC) of the restored images after tampering with known GANs is higher than 0.9 and the average NCC reaches around 0.8 with tampering by unknown GANs, which shows good robustness and generalization. In addition, the secret embedded information is well concealed and the average peak signal to noise ratio(PSNR) is about 30 dB.