基于动作单元强度曲线相似性的深度伪造人脸视频检测

Deep Fake Face Video Detection Based on Intensity Curve Similarity of Action Unit

通过分析现阶段换脸视频的合成效果和不同语者说话时独特的面部运动模式,提出一种的深度的造人脸视频的检测方法。实验分成训练和测试两部分,训练阶段包括面部动作单元的选择、Savitzky-Golay滤波器拟合去噪、分类器的训练等。测试阶段先使用Openface2.0提取待测视频中面部动作单元数据,然后截取有效讲话时间段的面部动作单元强度曲线,对曲线进行拟合,计算出同一语者的待测视频和样本视频中同类动作单元的曲线相似值,提供分类器检测,为选取代表性的动作单元,采用了随机森林法,并用消融实验验证其有效性,最后以ROC曲线评价该方法检测效果。在私有数据库中,该方法的检测效果较好,AUC值为0.953。研究表明,针对视频内容为人说话时的换脸视频,采用基于动作单元强度曲线相似值的检测方法,能有效检测出视频的真伪。

A detection method of deep fake face video was proposed by analyzing the synthesis effect of face-changing videos at the present stage and the unique facial movement patterns of different speakers.The experiment was divided into two parts:training and testing.The training stage included the selection of facial action units,the fitting and denoising of Savitzky-Golay filter and the training of classifiers.In the test stage,Openface2.0 was firstly used to extract the facial action unit data in the video to be tested.Then the intensity curve of facial action unit in the effective speech period was intercepted.After denoising,the curve similarity of the same action unit in the test video and sample video from the same speaker was calculated.Classifier detection was afforded to select the representative action units.Radom forest was adopted,and ablation experiments were used to verify its effectiveness.Finally,ROC curve was applied to evaluate the detection effect of this method.In the private database,the detection effect of this method was excellent,with an AUC value of 0.953.The research showed that the authenticity could be effectively detected by the detection method based on the intensity curve similarity value of the action unit for the face-changing videos of people talking.