Face Swapping Video Detection - Wang Yang, Junfeng Xiong, Liu Yan, Hao Xin, Wei Tao- DEF CON China 1

Recent developments of fabricating faces in videos such as Deepfakes have raised significant concerns that these deep learning techniques may be abused to create pornographic video or fake propaganda. In Deepfakes videos, the faces of a person are replaced with the faces of another one. And these faked videos are nearly indistinguishable for human. We find CNN-based networks can effectively distinguish DeepFakes videos from the real ones and present two effective methods. Firstly, we use a simple yet effective CNN architecture with several convolutional layers to build a powerful DeepFakes detector. Secondly, we find a FaceNet based method is an effective binary classifier. FaceNet is one of the state-of-the-art convolutional neural networks for face recognition, which could catch high-level features of faces. We use these features to train an SVM classifier. The two methods demonstrate successful detection reaching an accuracy rate of 99% and 94% respectively among our tests. Wang Yang is a senior security researcher of Baidu X-Lab. His interests lie in face recognition, adversarial learning, and data mining. He maintains and actively contributes to Advbox project that is an open source toolbox for AI safety. Junfeng Xiong(Jay Xiong) is an AI security researcher at Baidu X-Lab. His research interests cover deep learning security, privacy and IOT. Liu Yan (Dou Goodman), Head of AI security team of Baidu X-Lab, is a technology writer of AI Safety Trilogy. His research interests include AI and network security. He starts the open source project Advbox. Hao Xin has been engaged in security product development for many years in Baidu. His main research directions include object detection and image classification. Dr. Tao (Lenx) Wei is the head of Baidu X-Lab. Prior to joining Baidu, he was an associate professor at Peking University. His research interests include software analysis and system protection, web trust and privacy, programming languages, and mobile security.