面向浅层特征高频分量的深度伪造检测算法

Deepfake Detection Algorithm for High-Frequency Components of Shallow Features

近年来,深度伪造技术大幅提升了合成人脸的真实感,且相较于传统伪造方法,其生成的虚假视频更加难以分辨。基于深度伪造图像视觉伪影常常存在于特征提取网络浅层特征高频分量中这一特性,设计了一种面向浅层特征高频分量的深度伪造图像检测算法。针对高通滤波器的缺陷,本实验在拉普拉斯金字塔的基础上设计了一种具有更好的过滤性能的高频残差提取模块。在增强模块中,使用Convolutional Block Attention Module(CBAM)增加特征图关键区域以及关键特征通道的权重,提升特征图的空间以及通道相关性。针对深层网络中高频分量学习优先级低的问题,设计了一种图像梯度损失算法,防止高频信息随着网络的加深而丢失。将梯度中心化引入AdamW优化器,解决了深度伪造检测模型训练时间长、泛化性差的问题。所提两种模型在FaceForensics++和Celeb-DF数据集上的准确率均优于主流算法,证明了算法的有效性以及泛化性。

Deepfake techniques have dramatically improved the realism of synthetic faces in recent years.And the fake videos it generates are more difficult to distinguish than traditional forgery methods.Based on the characteristic that visual artifacts of depth forgery images often exist in the high frequency components of shallow features in feature extraction network,a detection algorithm for depth forgery images oriented to the high frequency components of shallow features is designed.First,a highfrequency residual extraction module based on Laplace’s pyramid with better filtering performance is designed to address highpass filters’shortcomings.Second,the Convolutional Block Attention Module(CBAM)is used to increase the weights of key regions of the feature map and key feature channels to improve the spatial and channel correlation of the feature map in the enhancement module.Then,an image gradient loss is designed to prevent the loss of highfrequency information as the network deepens to address the problem of low learning priority of highfrequency components in deep networks.Finally,gradientcentralization is introduced into the AdamW optimizer to solve the problems of long training time and poor generalization of deep forgery detection models.Two models proposed outperform mainstream algorithms in terms of accuracy when validated on the FaceForensics++and CelebDF datasets,demonstrating the algorithms’effectiveness and generalization.