基于改进ResNet34的深度人脸伪造检测方法研究  

Research on deep face forgery detection method based on improved ResNet34

作　　者：王雅坤 张宝林 王兆成 马琬雲 郭仕佳 WANG Yakun;ZHANG Baolin;WANG Zhaocheng;MA Wanyun;GUO Shijia(School of Electronics and Information Engineering,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学电子信息工程学院,天津300401

出　　处：《河北工业大学学报》2024年第6期44-51,共8页Journal of Hebei University of Technology

基　　金：河北省自然科学基金资助项目(F2023202013)。

摘　　要：深度人脸伪造技术的非法应用对人民群众的财产安全造成了严重危害,当前伪造视频检测准确率低且面对新型伪造技术的泛化能力差。针对以上问题,提出了一种注意力机制改进ResNet34的深度人脸伪造检测方法:引入了高效通道注意力模块,无需进行降维和升维的操作,从而保留了原始通道特征的信息完整性;将伪造图像中的伪影区域提取出来,作为主干特征输入模型,增强模型对人脸局部特征的检测性能;利用多尺度滑动窗口和不同的混合函数来生成带有伪影区域位置信息的标注,方便伪影检测模块对伪影特征的检测。实验结果表明,本研究方法效果显著,在FF++(c23)数据集上检测准确率为97.88%,AUC值为99.84%。The illegal application of deep face forgery technology has caused serious harm to the people’s property secu-rity,and the current forgery video detection accuracy is low and the generalization ability is poor in the face of new forg-ery technology.Aiming at solving the above problems,an attention mechanism is proposed to improve the deep face forg-ery detection method of ResNet34.First,an efficient channel attention module is introduced,which eliminates the need for dimensionality reduction and dimensionality upgrading,thus preserving the information integrity of the original chan-nel features;second,the artifact region in the forged image is extracted and input into the model as a backbone feature,which enhances the model’s performance of detecting local features of the face;third,the labeling with the position infor-mation of the artifact region is generated by using the multiscale sliding window and different mixing functions,which fa-cilitates the detection of artifact features by the artifact detection module.The experimental results show that the method of this study has a detection accuracy of 97.88%and an AUC value of 99.84%on the FF++(c23)dataset,and compared with the latest methods,the method of this study has the best generalization ability,which proves the validity of the meth-od proposed in this study.

关 键 词：深度人脸伪造 ResNet34 高效通道注意力 伪影检测 多尺度滑窗 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]