摘要
针对现有的深度伪造检测方法缺少关注图像空间域信息及模型复杂度较高的问题,提出一种使用空间域特征增强的通道自适应深度伪造图像检测方法.首先提取图像空间域信息并归一化空域特征图,将空域信息作为第4个通道传入网络;其次,在主干网络中加入SE-Layer模块,对4个通道的权重进行重建,解决通道间的异构性问题;最后,设计了半自动的预训练策略,进一步提升了模型训练效率和准确率.以深度伪造人脸检测为例,在7个不同来源的数据集上进行检测实验.结果表明,即使在未使用数据增强技术的情况下,文中方法优于基线方法,在效果最差的StyleGAN2数据集上准确率也达到99.60%,AP指标达到98.00%.
Aiming at the problems that the current deepfake detection method lacks attention to the spatial domain information of the image and the high complexity of the model,a channel adaptive deepfake image detection method using spatial domain feature enhancement is proposed.Firstly,spatial domain features of the images are extracted and normalized,which will be passed into the network as the fourth channel;secondly,SE-Layer module is incorporated into the backbone network to reconstruct the weights of the four channels,effectively addressing the heterogeneity issue between channels;finally,a semi-automatic pre-training strategy is designed to further improve the training efficiency and accuracy of the model.Taking deepfake face detection as an example,the detection experiment is conducted on seven datasets from different sources.The results show that even without data augmentation,the method in this article is better than the baseline method,with an accuracy of 99.60%and an AP index of 98.00%on the StyleGAN2 dataset with the worst performance.
作者
李佳林
沈哲
Li Jialin;Shen Zhe(Northeastern University at Qinhuangdao,Qinhuangdao 066004)
出处
《计算机辅助设计与图形学学报》
北大核心
2025年第2期313-320,共8页
Journal of Computer-Aided Design & Computer Graphics
关键词
深度伪造检测
空间域特征
通道自适应网络
预训练
deepfake detection
spatial domain feature
channel adaptive network
pre-training
作者简介
李佳林(2002-),男,在校学生;通信作者:沈哲(1982-),男,博士,讲师,CCF会员,主要研究方向为计算机图形学、虚拟现实与增强现实技术.shenzhe@neuq.edu.cn。
