[techweb] on August 19, recently, Ali security Turing laboratory announced that it had successfully developed deepfake detection technology for face changing videos of multiple people, and the papers on this technology were included in acmmm2020, an International Academic Summit. < / P > < p > deepfake detection technology has the value of many practical application scenarios. For example, when an attacker changes the face of the main character of an indecent video into a target face for transmission, deepfake detection technology can “identify the fake and seek the truth” and trace the truth. < / P > < p > at present, the detection schemes for deepfake face changing video are mainly divided into two categories: frame level detection and video level detection. Frame level detection method has high labeling cost, and it also needs to convert frame level prediction into video level prediction. It has high requirements for fusion technology, and it is easy to lead to missed detection or false detection. Too much research on video level detection focuses on detection model constructed according to time sequence, which results in detection effect being limited. < p > < p > the algorithm engineer of Ali security Turing laboratory introduced to Xi that in order to better detect some tampered deepfake videos, Ali security Turing laboratory proposed a new detection method. This method is simple to label, and can help neural network to better learn face features and achieve better detection effect. < p > < p > Ali security Turing laboratory also found that the attacker tampered with the video. Because the attacker tampered with the video in a single frame, the same face would have some jitter in the adjacent frames. Therefore, the researchers designed a new detection module to detect these jitters and assist the recognition. < / P > < p > in addition, the detection methods previously proposed by the industry are mostly applicable to single video face tampering or all face tampering in multi person video. Alibaba security constructs a partial attack data set, which makes up for the blank of deepfake detection data set in the scene of only one or several face tampering in multi face video. < p > < p > through technical innovation, the deepfake detection technology of Ali security Turing laboratory ranks first in the field of video level detection and frame level detection. Wang Shuhui, CO researcher of the technology and associate researcher of the Institute of computing, Chinese Academy of Sciences, believes that the research results also have important enlightening significance for the research of general video multi instance learning and annotation technology. < p > < p > in March this year, Alibaba released a new generation of security architecture, which is committed to preventing security threats from the source, building a security system, and building a digital infrastructure security model room. As the core AI technology of the new generation security architecture, the deepfake detection technology developed by Alibaba security plays an important role in the security construction of digital infrastructure and successfully realizes its application. < p > < p > Huatang, a senior algorithm expert at Ali security Turing laboratory, said that so far, Alibaba has used the detection technology in content security scenarios, and the layout will be carried out in the live broadcast scene in the future.