GRU和Involution改进的深度伪造视频检测方法  

Deepfake Video Detection Method Improved by GRU and Involution

作　　者：刘亚琳 芦天亮[1] LIU Yalin;LU Tianliang(College of Information and Cyber Security, People’s Public Security University of China, Beijing 100038, China)

机构地区：[1]中国人民公安大学 信息网络安全学院,北京 100038

出　　处：《计算机工程与应用》2023年第22期276-283,共8页Computer Engineering and Applications

基　　金：国家社会科学基金重大项目(21&ZD193);中国人民公安大学基本科研业务费重大项目(2020JKF101)。

摘　　要：近年来深度伪造视频在网络上广泛传播,造成了负面影响。针对现有检测模型准确率低和信息提取不够充分和全面的问题,提出了一种GRU(gated recurrent unit)和Involution改进的深度伪造视频检测方法。首先使用VGG19作为主干网络提取空间特征,并将Involution算子嵌入主干网络,从空间和通道信息两方面加强了人脸图像的空间建模能力。然后通过主胶囊层关注特征的位置信息和使用GRU提取帧间的时序特征。最后在训练模型阶段使用focalloss作为损失函数来平衡样本。在Deepfakes、FaceSwap和Celeb-DF数据集中进行测试,实验结果表明该方法优于主流检测方法,改进对比实验进一步证明了检测方法的有效性。In recent years,the wide spread of deepfake video on the network has caused a negative impact.In order to solve the problems of low accuracy of existing detection models and insufficient and comprehensive information extraction,an improved deepfake video detection method based on gated recurrent unit(GRU)and Involution is proposed.Firstly,a feature extraction network is constructed based on Involution operator to extract global feature information,which enhances the spatial modeling ability of face image from spatial and channel information.Then,the temporal features are extracted through the location and inter-frame information of the main capsule layer and GRU concern features.Finally,focalloss is used as the loss function to balance the samples in the training model phase.The method is tested in Deepfakes,FaceSwap and Celeb-DF datasets,and the experimental results show that the method is better than the mainstream detection methods.The improved comparative experiments further prove the effectiveness of the detection method.

关 键 词：深度伪造 门控循环单元(GRU) INVOLUTION 胶囊网络 focalloss 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]