【摘要】 针对现有生成对抗网络(GAN)生成人脸反取证方法攻击迁移性不强的问题，提出了一个基于图像小波域自适应干扰的GAN生成人脸反取证方法以提升攻击迁移性.该方法通过对GAN生成人脸图像的小波域信息(即图像经过小波分解后的频率分量)施加扰动以实现其对取证模型的抵抗，并且分别在空域和频域上基于最小可觉察误差(JND)设计自适应扰动阈值，对图像不同像素点位置设置不同的扰动强度，从而增强扰动的人眼不可感知性.此外，还设计了一种数据增强方式对反取证人脸进行数据分布多样性扩充，以进一步提升攻击迁移性.实验结果表明，与6种基线方法相比，所提方法生成的反取证人脸在保证扰动对人眼不可感知前提下具有更强的攻击迁移性.更多还原

【Abstract】 Aiming at the insufficient attack transferability of existing generative adversarial network(GAN)-generated face anti-forensics methods, a GAN-generated face anti-forensics method based on image wavelet domain adaptive perturbations is proposed to improve the attack transferability. The proposed method resists the forensic models by adding perturbations to the wavelet domain information of GAN-generated facial images, which are the frequency components after the image wavelet decomposition. Furthermore, adaptive perturbation thresholds are designed based on just noticeable distortion(JND) in both the spatial and frequency domains, setting different perturbation strengths for different pixel positions in the image, and thereby enhancing the imperceptibility of the perturbations to the human eyes. In addition, a data argumentation approach is designed to expand the distribution diversity of the anti-forensics image, thereby further improving the attack transferability. Experimental results show that compared with the six baseline methods, the anti-forensics image generated by the proposed method can achieve stronger attack transferability while ensuring the perturbation imperceptibility to the human eyes.更多还原