基于空洞卷积和注意力机制的深度伪造检测

Deepfake detection based on dilated convolution and attention mechanism

人脸深度伪造图像的生成过程是对目标人脸进行替换,由于不同图片的肤色和光照等差异,在人脸周围会产生比较分散的伪影边界,且对人脸进行裁剪之后会包含部分背景信息的干扰。针对深度伪造图像的生成特点,区别于常规的CNN分类识别网络,设计了一种符合伪造痕迹特点的检测网络。首先,利用Resnet34卷积神经网络对待检测的图像提取初步的伪造特征;其次,经过空洞卷积提高网络模型的感受野来更好地捕获比较分散的伪造痕迹;再由空间注意力模块进行权重的重加权,减少其他背景信息的干扰;最后使用多层全连接神经网络对特征进行分类,达到对深度伪造图片进行准确分类的目的。在Faceforensics++、Celeb-DF、DFDC三个主流数据集上进行实验,并取得比当前其他方法更好的效果。实验结果表明,所提方法结合了空洞卷积和注意力机制,应对不同人脸截取比例的伪造图像拥有更好的适应性。

The generation process of the face deepfake image is to replace the objective face.Due to the differences in skin color and illumination in different pictures,relatively scattered artifact boundaries will be generated around the face,and part of the background information will be left after the face is cropped.According to the generation characteristics of deepfake images,a detection network that meets the characteristics of forged traces is designed,which is different from the conventional convolutional neural network(CNN)classification and recognition network.The Resnet34 CNN is used to extract the preliminary fake features of the image under detection.The dilated convolution is used to improve the receptive field of the network model to better capture the scattered fake traces.The spatial attention module is used to reweight the weight to reduce the interference of other background information.Finally,a multi-layer fully-connected neural network is used to classify the features to achieve the purpose of accurately classifying the deepfake images.Experiments were conducted on the three mainstream data sets of Faceforensics++,Celeb-DF and DFDC,and the effect was better than that of the other current methods.The experiment results show that the method combines the dilated convolution and attention mechanism,and has better adaptability for deepfake images with different face interception proportions.