摘要： 人脸深度伪造检测技术对于打击虚假图像/视频泛滥具有至关重要的意义。提出了一种融合传统特征与神经网络的检测算法，算法结合了传统特征具有可解释性与神经网络高准确率的优点，利用图像灰度共生矩阵以及XceptionNet组成双特征提取模块，然后在全卷积网络中充分考虑双流融合特征信息，最终根据网络多损失实现图像真伪分类判决。在FaceForensics++数据集上进行了训练和测试，实验结果表明，相比现有深度学习算法，检测准确率有明显提升。而且由于引入的纹理特征具有一定的可解释性，表现出良好的鉴别性能。

中图分类号： TP181
文献标识码： A
DOI： 10.19358/j.issn.2096-5133.2021.02.006
引用格式： 杨雨鑫，周欣，熊淑华，等. 融合传统特征与神经网络的深度伪造检测算法[J].信息技术与网络安全，2021，40(2)：33-38，44.

Research on deepfakes detection combining traditional features and neural network

Abstract： DeepFakes detection is significant to combat the spread of forgery video. Aiming at the task of deepfakes detection, a method combining traditional features and neural network is proposed. The method combines the interpretability of traditional features and high accuracy of the neural network. This paper used the gray level co-occurrence matrix and XceptionNet to form two feature extraction modules, then learned the dual-stream fusion feature information in the fully convolutional network. The image was distinguished according to multiple losses in the network finally. Our method was tested over benchmarks of the FaceForensics++ datasets. The experimental results show that compared with the state-of-the-art deep learning algorithms, the detection accuracy has been significantly improved. It shows promising discrimination performance due to the introduction of texture feature interpretability.

Key words : deepfakes；image forensics；feature fusion；gray level co-occurrence matrix(GLCM)；convolutional neural network(CNN)