网络与信息安全学报 ›› 2021, Vol. 7 ›› Issue (2): 151-160.doi: 10.11959/j.issn.2096-109x.2021032

• 学术论文 • 上一篇    下一篇

基于双层注意力的Deepfake换脸检测

龚晓娟1,2, 黄添强1,2, 翁彬1,2, 叶锋1,2, 徐超1,2, 游立军3   

  1. 1 福建师范大学数学与信息学院,福建 福州 350117
    2 数字福建大数据安全技术研究所,福建 福州 350117
    3 福建省灾害天气重点实验室,福建 福州 350001
  • 修回日期:2020-09-29 出版日期:2021-04-15 发布日期:2021-04-01
  • 作者简介:龚晓娟(1995- ),女,福建福州人,福建师范大学硕士生,主要研究方向为数字多媒体取证。
    黄添强(1971- ),男,福建仙游人,博士,福建师范大学教授、博士生导师,主要研究方向为机器学习、数字多媒体取证。
    翁彬(1981- ),男,福建福州人,博士,福建师范大学讲师,主要研究方向为机器学习及应用。
    叶锋(1978- ),男,福建福州人,博士,福建师范大学副教授,主要研究方向为多媒体信号处理、计算机视觉。
    徐超(1981- ),男,湖北天门人,福建师范大学讲师,主要研究方向为视频篡改检测。
    游立军(1974- ),男,福建莆田人,福建省灾害天气重点实验室高级工程师,主要研究方向为气候数据分析。
  • 基金资助:
    国家重点研发计划专项基金(2018YFC1505805);国家自然科学基金(62072106);国家自然科学基金(61070062);应用数学福建省高校重点实验室(莆田学院)开放课题(SX201803)

Deepfake swapped face detection based on double attention

Xiaojuan GONG1,2, Tianqiang HUANG1,2, Bin WENG1,2, Feng YE1,2, Chao XU1,2, Lijun YOU3   

  1. 1 College of Mathematics and Informatics, Fujian Normal University, Fuzhou 350117, China
    2 Digital Fujian Institute of Big Data Security Technology, Fuzhou 350117, China
    3 Fujian Key Laboratory of Severe Weather, Fuzhou 350001, China
  • Revised:2020-09-29 Online:2021-04-15 Published:2021-04-01
  • Supported by:
    The National Key Program for Developing Basic Science(2018YFC1505805);The National Natural Science Foundation of China(62072106);The National Natural Science Foundation of China(61070062);Key Laboratory of Applied Mathematics of Fujian Province University (Putian University)(SX201803)

摘要:

针对现有Deepfake检测算法中普遍存在的准确率低、可解释性差等问题,提出融合双层注意力的神经网络模型,该模型利用通道注意力捕获假脸的异常特征,并结合空间注意力聚焦异常特征的位置,充分学习假脸异常部分的上下文语义信息,从而提升换脸检测的有效性和准确性。并以热力图的形式有效地展示了真假脸的决策区域,使换脸检测模型具备一定程度的解释性。在 FaceForensics++开源数据集上的实验表明,所提方法的检测精度优于MesoInception、Capsule-Forensics和XceptionNet检测方法。

关键词: Deepfake, 换脸检测, 假脸检测, 注意力

Abstract:

In view of the existing Deepfake detection algorithms, such problems as low accuracy and poor interpretability are common.A neural network model combining the double attention was proposed, which used channel attention to capture the abnormal features of false faces and combined the location of spatial attention to focus the abnormal features.To fully learn the contextual semantic information of the abnormal part of the false face, so as to improve the effectiveness and accuracy of face changing detection.In addition, the decision-making area of real and fake faces was shown effectively in the form of thermal diagram, which provided a certain degree of explanation for the face exchange detection model.Experiments on FaceForensics ++ open source data set show that the detection accuracy of proposed method is superior to MesoInception, Capsule-Forensics and XceptionNet.

Key words: Deepfake, face swap detection, fake face detection, attention

中图分类号: 

No Suggested Reading articles found!