One second to see through the "AI face-changing" star face-changing, face-changing fraud nowhere to hide.

Everyone has been exposed to AI face-changing technology to a certain extent, and the face-changing App“ZAO "became popular overnight, because uploading a photo can transform you into a character in the play, and you can compete with Tony Leung Chiu Wai, Andy Lau, Yang Mi and other stars on the same stage, but it is also defeated overnight because of the risk of revealing facial information. Not long ago, the special effect of "Ant Hey" to make still photos move became popular on short video platforms.

These online celebrity applications are based on Deepfake technology. Although they are also used in scenes such as later movies and virtual idols, the negative threat of this technology is magnified.

How to tell whether the face in the video is true or not? At the 202 1 World Artificial Intelligence Conference, the artifact of "AI changing face" in one second-DeepReal deep forgery content detection platform was unveiled.

The martial arts in the world are fast and unbreakable. Within 30 milliseconds, a frame can be detected. In a second, you can recognize a fake face.

This is the nemesis of the DeepReal deep forgery content detection platform "AI Face Changing" developed by Beijing Ruilai Intelligent Technology Co., Ltd. (hereinafter referred to as Ruilai), the incubator of Tsinghua University Institute of Artificial Intelligence.

How high is its detection accuracy? The test results show that the accuracy of DeepReal in academic data sets and network data generated by mainstream methods such as ZAO has reached 99%, far exceeding the best results announced by Facebook in the previous Deepfake Detection Challenge.

The data under DeepReal has reached tens of millions, and the data set has covered three categories: academic depth pseudo-data set, network depth pseudo-data set and self-developed depth pseudo-data set. By combining Bayesian learning framework and deep neural network, the uncertainty of the model in predicting new samples is estimated.

Tang Jiahua, vice president of RealAI, believes that DeepReal can be quickly applied in such scenes as false content screening on the Internet, face data auditing, and authenticity verification of image evidence.

On the legal level, in 2020, it is forbidden to use information technology to forge other people's portrait rights and voices, which has been written into the civil code, putting a "curse" on deep forgery technology. Then, on the technical level, how to continuously develop high-quality detection and cracking tools?

Setting standards is the first step.

RealAI and the National Center for Industry and Information Security jointly formulated the first domestic standard related to face model algorithm security, namely "Technical Specification for Information Security Technology Face Comparison Model Security". RealAI and ICT Institute, as the second compiling unit, jointly compiled the Blue Book of Artificial Intelligence Security Framework (2020) under the guidance of the Network Security Bureau of the Ministry of Industry and Information Technology. RealAI is currently involved in the formulation of national standards related to machine learning algorithm security, as well as industry standards related to artificial intelligence synthetic audio and video and artificial intelligence biometrics.

More and more organizations are involved in the detection and standard setting of AI. "We are formulating various standards for artificial intelligence. This year, Shanghai has approved the establishment of the local standard "Application Specification for Classification and Classification of Face Recognition in Public Places". Chen Mingang, executive director of the Institute of Artificial Intelligence Governance of Shanghai Computer Software Technology Development Center, told the IT Times reporter that face recognition will become the highland of ai standards and detection tools.