深度学习的迅速发展，显著提高了面部交换图像生成的质量和效率。利用Deepfake等生成的面部交换，无论是人工检测还是自动检测都很难分辨真伪。面部交换技术既可用于积极的用途，也有可能用于非法的面部伪造。因此研究最先进的面部交换技术，既是正面发展的需要，也是应对伪造滥用风险，探索伪造检测技术的需要。本文针对面部交换技术和伪造检测技术两个方面展开研究，主要内容和研究成果如下：

针对现有面部交换算法面对复杂姿态或光照时伪造痕迹明显的问题，提出了一种基于三维重建的面部交换方法。首先，训练了一个端对端的位置映射图回归网络，用于实现单张图片的面部三维重建。然后，对要进行面部交换的目标图像进行三维重建，根据重建得到的三维形状，获得相应的颜色空间和顶点。最后，结合源图的颜色空间和目标图的顶点信息，渲染获得最终的面部交换图像。实验证明了基于三维重建的面部图像交换方法的有效性和鲁棒性。特别是与利用二维人脸图像进行面部交换的结果相比，该方法在复杂姿态或光照场景下生成的图像更加真实自然。

针对大多数面部伪造检测算法模型复杂、训练周期长等问题，提出了一种轻量级的深度学习模型，用于检测面部伪造图像。首先，分析了Deepfake算法的生成原理，由于Deepfake只能合成固定分辨率的面部图像，导致作为前景的面部区域与作为背景的面部区域，两者的图像压缩率不同，两者的拼接会留下明显的伪造痕迹。然后，使用误差水平分析方法来突出拼接图像的压缩率差异，生成的中间图保留了压缩率不一致的关键特征，去除了与篡改无关的大量特征。最后，构建一个二分类的深度学习模型，用于检测该图像是否是伪造。实验表明，基于该方法的卷积神经网络架构，对于面部伪造图像的检测可以达到97%以上的准确率。同VGG16，ResNet50，ResNet101和ResNet152四种网络模型相比，该方法的模型不仅轻量化，易于训练，而且实现了更高的效率。

The quality and efficiency of generating face-swap images have been markedly strengthened by deep learning. The face-swap manipulations by DeepFake are so real that it is tricky to distinguish authenticity through automatic or manual detection. Face-swap technology can be used for both positive and illegal purposes. Therefore, the study of the most advanced face-swap technology is not only the need of positive technology development, but also the need of resisting and detecting face forgery. In this paper, we focus on the face-swap technology and forgery detection technology. The main works are as follows:

In order to solve the problem of obvious forgery trace in complex pose or illumination in existing face-swap algorithms, a face-swap method based on 3D reconstruction is proposed. First, an end-to-end position map regression network is trained to realize the facial 3D reconstruction of a single image. Then, the target 3D image is generated by the 3D reconstruction network. According to the reconstructed 3D shape, the corresponding color space and vertices are obtained. Finally, combining the color space of the source image and the vertex information of the target image, the final face is rendered. Experiments show that the face-swap method based on 3D reconstruction is effective and robust. Especially, compared with the result of face-swap using 2D face image, the images generated by this method are more realistic and natural in various complex pose or illumination scene.

In view of the complexity and long training period of most facial forgery detection algorithms, a lightweight deep learning model is proposed to detect face-swap forgery images. First, the generation principle of deepfake algorithm is analyzed. Because deepfake can only generate limited resolution, this algorithm results in two different image compression ratios between the fake face area as the foreground and the original area as the background, which would leave distinctive counterfeit traces. Then, the error level analysis method is used to highlight the compression ratio difference of the stitching image. Finally, a binary deep learning model is constructed to detect whether the image is a forgery or not. The experimental results show that the detection accuracy of the face forgery image based on the convolutional neural network architecture can reach more than 97%. Compared with the network models of vgg16, resnet50, resnet101 and resnet152, the model of this method is not only lightweight, easy to train, but also more efficient. Specifically, without loss of accuracy, the amount of computation can be significantly reduced.