摘要
近年来,深度学习的快速进步使得经过多媒体篡改的人脸视频达到了以假乱真的程度。这种用深度学习的架构来生成虚假人脸的方法被称为“Deepfake”。现有的Deepfake检测方法应用于高分辨率的原视频时性能尚可,然而应用于经过高度压缩的低质量视频时表现欠佳。针对大部分现有算法对视频的帧间信息利用不够充分这一问题,本文提出了一种基于时空信息的检测算法。首先,该方法设计了一种时空注意力架构,同时提取了空间和时间上的注意力以抑制无关信息;然后,对提取出的时空注意力权重信息通过改进深度可分离卷积(Xception)和卷积门控循环网络(ConvGRU)进行深层次的提取,ConvGRU用于获取在改进Xception网络降维前丢失的帧间信息;最后,使用判别器进行二分类,通过实验训练模型,在FF++数据集低质量视频上获得了97%的准确率,取得了良好的效果。
In recent years,the rapid progress of deep learning has made it difficult for people to distinguish between real videos and multimedia tampered face videos.The method of generating fake faces using a deep learning architecture is called"Deepfake".The existing Deepfake detection methods perform well when applied to high-resolution original videos.However,their detection performance is unsatisfactory for low quality videos that have been highly compressed.To address the issue of insufficient utilization of inter frame information in most existing algorithms,this paper proposes a detection algorithm based on dual stream attention.Firstly,this method designs a dual-branch attention architecture that extracts both spatial and temporal attention to suppress irrelevant information.Then,the extracted spatiotemporal attention weight information is extracted at a deeper level through improved deep separable convolution(Xception)and Convolutional Gated Recurrent Network(ConvGRU).ConvGRU aims to obtain inter frame information lost before the improved Xception network dimensionality reduction.Finally,using a discriminator for binary classification,the model is trained through experiments and achieves an accuracy of 97%on low quality FF++datasets,achieving good results.
作者
黄夏馨
HUANG Xiaxin(College of Electronic Information,Soochow University,Suzhou 215000,Jiangsu,China)
出处
《智能计算机与应用》
2024年第10期99-106,共8页
Intelligent Computer and Applications
关键词
虚假人脸检测
双流注意力
特征处理
deepfake detection
dual-branch attention
feature processing
