基于深度学习的多模态人脸篡改识别
Multimodal Face Tampering Recognition Based on Deep Learning
李杰1
作者信息
摘要
深度伪造技术是一种基于深度学习的图像和音视频合成技术.它使用深度神经网络生成高度逼真的虚构内容,尤其在人脸篡改方面,已对日常生活产生了不良影响.随着深度学习的发展,传统的卷积神经网络难以准确识别当前的人脸篡改行为.为应对该挑战,提出了一种多模态人脸篡改识别方法,创新性地结合了频域处理、错误水平分析以及语义信息进行检测.经过在FaceForensics++数据集上进行实验测试,结果表明,该方法的准确率高达 83.10%,是一种有效的人脸篡改检测方法.
Abstract
Deepfake technology is a deep learning based image and audio video synthesis technique.It uses deep neural networks to generate highly realistic fictional content,especially in the area of facial tampering,which has had a negative impact on daily life.With the development of deep learning,traditional convolutional neural networks are difficult to accurately recognize current facial tampering behavior.To address this challenge,a multimodal facial tamper recognition method has been proposed,which innovatively combines frequency domain processing,error level analysis,and semantic information for detection.After experimental testing on the FaceForensics++dataset,the results show that the accuracy of this method is as high as 83.10%,making it an effective face tamper detection method.
关键词
深度伪造/人脸篡改/多模态/离散余弦变换/深度学习Key words
deepfakes/face tampering/multimodal/discrete cosine transformation/deep learning引用本文复制引用
基金项目
郑州大学2023年大学生创新创业训练计划资助项目(202310459110)
出版年
2024
NETL
NSTL
万方数据