A Novel Deep Learning Approach for Deepfake Image Detection
A Novel Deep Learning Approach for Deepfake Image Detection
复制标题
用于 Deepfake 图像检测的新型深度学习方法
作者：
Ali Raza;Kashif Munir;M. Almutairi
Deepfake是在合成媒体中利用一个人现有的媒体来生成虚假的视听内容。deepfake用假媒体代替人的脸和声音，使其看起来更逼真。虚假媒体内容的产生是不道德的，是对社会的威胁。如今，深度造假被高度滥用于网络犯罪中，用于身份盗窃、网络勒索、假新闻、金融欺诈、名人假淫秽视频勒索等等。根据Sensity最近的一份报告，超过96%的深度造假是淫秽内容，大多数受害者来自英国、美国、加拿大、印度和韩国。2019年，网络犯罪分子制作了一名首席执行官的虚假音频内容，打电话给他的组织，要求他们向他们的银行账户转账24.3万美元。深度造假犯罪每天都在上升。深度假媒体检测是数字取证领域的一大挑战和高要求。必须建立一种先进的研究方法，通过检测深度虚假内容来保护受害者免受勒索。我们研究的主要目的是使用一个有效的框架来检测深度假媒体。本文提出了一种基于VGG16和卷积神经网络结构混合的深度假预测器（DFP）方法。利用基于真假人脸的深度伪造数据集构建神经网络技术。exception、NAS-Net、Mobile Net和VGG16是比较中使用的迁移学习技术。所提出的DFP方法在深度伪造检测中达到了95%的精密度和94%的准确度。我们提出的新颖DFP方法优于迁移学习技术和其他最先进的研究。我们新颖的研究方法可以帮助网络安全专业人员通过准确检测深度假内容来克服与深度假相关的网络犯罪，并使深度假受害者免受勒索。
Deepfake is utilized in synthetic media to generate fake visual and audio content based on a person’s existing media. The deepfake replaces a person’s face and voice with fake media to make it realistic-looking. Fake media content generation is unethical and a threat to the community. Nowadays, deepfakes are highly misused in cybercrimes for identity theft, cyber extortion, fake news, financial fraud, celebrity fake obscenity videos for blackmailing, and many more. According to a recent Sensity report, over 96% of the deepfakes are of obscene content, with most victims being from the United Kingdom, United States, Canada, India, and South Korea. In 2019, cybercriminals generated fake audio content of a chief executive officer to call his organization and ask them to transfer $243,000 to their bank account. Deepfake crimes are rising daily. Deepfake media detection is a big challenge and has high demand in digital forensics. An advanced research approach must be built to protect the victims from blackmailing by detecting deepfake content. The primary aim of our research study is to detect deepfake media using an efficient framework. A novel deepfake predictor (DFP) approach based on a hybrid of VGG16 and convolutional neural network architecture is proposed in this study. The deepfake dataset based on real and fake faces is utilized for building neural network techniques. The Xception, NAS-Net, Mobile Net, and VGG16 are the transfer learning techniques employed in comparison. The proposed DFP approach achieved 95% precision and 94% accuracy for deepfake detection. Our novel proposed DFP approach outperformed transfer learning techniques and other state-of-the-art studies. Our novel research approach helps cybersecurity professionals overcome deepfake-related cybercrimes by accurately detecting the deepfake content and saving the deepfake victims from blackmailing.
DOI：
10.1109/wifs49906.2020.9360897
发表时间：
2020-12
期刊：
2020 IEEE International Workshop on Information Forensics and Security (WIFS)
影响因子：
--
作者：
Yassine Yousfi;Jan Butora;Eugene Khvedchenya;J. Fridrich
通讯作者：
Yassine Yousfi;Jan Butora;Eugene Khvedchenya;J. Fridrich