| 摘要: |
| 随着深度学习技术的飞速发展，以Deepfakes为代表的深度伪造技术开始充斥在互联网上的各个角落。Deepfakes借助于生成对抗网络和自动编码器技术，能够轻松替换人脸以及篡改人的表情信息。此类Deepfakes假视频可以制作虚假色情影片、谣言，传播假新闻，甚至影响政治选举，带来的社会影响极其恶劣。然而，针对此类伪造视频的检测技术还远远落后于生成技术，已有的工作都存在一定的局限性，并不能较好地对Deepfakes视频进行检测。本文首先对现有生成和检测工作进行综述，并分析了现有工作的缺陷，然后提出了基于EfficientNet的双流网络检测框架。通过在大规模开源数据集FaceForensics++测试，我们的检测技术可以在检测Deepfakes类假视频上平均准确率达到99%以上，并一定程度提高模型对抗压缩的能力。 |
| 关键词: 深度学习 深度伪造 检测 双流网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2020.02.07 |
| Received:January 16, 2020Revised:March 09, 2020 |
| 基金项目:本课题得到阿里实人认证项目资助。 |
|
| A Deepfakes detection technique based on two-stream network |
| LI Xurong,YU Kun |
| Alibaba Group, Hangzhou 310027, China |
| Abstract: |
| With the rapid development of deep learning technology, Deep forgery techniques, such as Deepfakes, are beginning to fill every corner of the Internet. By utilizing the generative adversarial networks and auto-encoder technology, the Deepfakes replace faces and tamper with facial expressions easily. The Deepfakes can produce fake pornography, spread rumors, spread fake news, and even influence political elections, leading to disastrous social consequences. However, the detection technology for this kind of fake videos is still far behind the generation technology, and the existing works have some limitations. This paper first summarizes the existing generation and detection works, and analyzes the defects of the existing works, then we propose the two-stream network detection framework based on the EfficientNet. By testing on a large open source dataset, FaceForensics++, our detection method was able to detect fake videos with an average accuracy of over 99%, and improve the ability of the model to resist compression to a certain extent. |
| Key words: deep learning deepfakes detection two-stream networks |
