Why is it said that AI face-changing is testing the edge of illegality?

If you are a person who often surfs the Internet, I believe you must have seen a video of "Yang Mi version of Huang Rong" - netizens used AI face-changing technology to replace Yang Mi's face with the face of Zhu Yin, who played Huang Rong in the 1994 version of "The Legend of the Condor Heroes". The expression of "Yang Mi version of Huang Rong" is lively and very realistic. Once released, it received 120 million clicks and 28,000 discussions on the Internet.

With a quick search, you will find quite a few similar videos and pictures: for example, the Chinese actor Xu Jinjiang looks similar to Thor in the currently hit Marvel movie "Avengers 4", and is often used as the protagonist in face-changing videos. The "Aquaman" in another movie "Aquaman" also joined the "face-changing army" because of his similar appearance. Netizens exclaimed: Thor, Aquaman, Xu Jinjiang, it's hard to tell them apart!

▲"Thor and Aquaman Xu Jinjiang" made by netizens

Daniel Wu swapped his face with Ni Dahong, the male and female protagonists of "The Temptation of Going Home" swapped faces, and even female anchors changed their faces online - you can see Liu Yifei, Tang Yan, and Fan Bingbing singing with guitars in the same live broadcast room...

So what exactly is this so-called AI face-changing technology?

AI face-changing is not as complicated as you think

The first face-changing video that caused a stir was a pornographic video of "Wonder Woman" Gal Gadot posted on Reddit by a netizen named DeepFakes in December 2017. He then posted indecent videos of several actresses, which caused collective protests from the actresses. Reddit subsequently blocked DeepFakes' account. After the farce came to an end, DeepFakes directly open-sourced the code of its AI face-changing project, and face-changing technologies derived from DeepFakes also sprang up like mushrooms after rain.

DeepFakes refers to replacing one face with another in an image or video. In fact, face swapping technology has long been used in the field of film production. For example, when actor Paul Walker died in an accident during the filming of "Fast and Furious 7", the crew used face swapping technology to simulate his face and completed the filming of the second half of the movie.

▲Paul Walker's face in "Fast and Furious 7" was synthesized by CGI technology

However, in the field of professional movies, face swapping is quite complicated, and professional video editors and CGI experts need to spend a lot of time and effort to complete face swapping in videos.

DeepFakes is actually an image processing technology based on deep learning. This technology is a new breakthrough in "face-changing". Even if you don't have artificial intelligence expertise and video production skills, you can complete an AI face-changing. In January 2018, FakeApp for Windows was released, and users can perform customized face-changing. The equipment and materials used are just a personal computer with an Nvidia GPU, 3~4GB of storage space, and materials for training neural networks.

So how does DeepFakes achieve face-changing?

The Studio51 team that produced the video of Leslie Cheung's face-changing face introduced in an article: The so-called face-changing technology is to use the generative adversarial network (GAN) to make two machine learning models compete. One ML model is trained on the data set and then creates a fake video, while the other model tries to detect the fake, until the other ML model cannot detect the fake, then the training can be ended.

AI face-changing is also an application of face recognition technology. To achieve "face-changing", the machine must first learn to recognize faces. The first thing AI does is recognition, and the second thing is face-changing. So it is not difficult to understand why the protagonists of many "face-changing" videos are celebrities we are familiar with, because the machine needs enough materials and data to improve the accuracy of recognition, and celebrities who are exposed in front of the camera in large quantities are natural application materials.

A mixed bag of people, testing the edge of illegality

The cost of face-changing technology is very low. A blogger pointed out that you only need to write a program script of about 200 lines to automatically replace the face in one picture with the face in another picture.

A very low-cost technology can achieve "face-swapping", which has triggered some negative discussions among netizens about the technology - will actors be divided into two types in the future, those who are good at acting and those who are good at acting? How much authenticity can be achieved in the production of film and television works? If face-swapping technology is used to forge and produce indecent videos, wouldn't it be easy for public figures to fall for it? "Fake videos" will intensify the spread of false news, which will greatly damage the credibility of videos as evidence.

The netizen Face Changing Brother who posted the video of Yang Mi's face being replaced by Athena Chu also directly removed the relevant video after doubts arose, and issued a statement saying that he hopes everyone will focus on the technology itself.

On April 1 this year, QuantumLiu, a technical staff of Studio51, a domestic visual effects company, released a 5-minute video online: the boy in the video "became" Leslie Cheung through AI face-changing technology, and sang Leslie Cheung's classic song "Thousands of Songs" during a live broadcast. This video triggered the emotion of countless fans, and then the producer of the video issued a statement: There are two main wishes for releasing this video - first, to share it with Leslie fans, and use this way to commemorate Leslie's unparalleled elegance; second, I also hope that everyone can see that AI face-changing technology can be used for more noble purposes.

The words of these professional technicians do reflect a current situation: if low-cost technology is used by people with ulterior motives, it is likely to have adverse effects. The immature technology of AI face-changing has already brought some negative impressions to the public.

People may be very familiar with "Photoshop" now. When seeing photos that don't conform to common sense, their first reaction is "Photoshop, right?" But many years ago, picture synthesis also caused people's "panic" just like today's AI face-changing. Today, the authority of videos is being challenged. The development of synthesis technology has led to the collapse of the public's trust in the media. If everything can be faked, what can we believe in the future?

In the information technology era, security issues are always present, and face recognition technology is actually facing security threats. There is no absolutely safe system, and the contradiction between function and defense always exists.

The public has doubts about technologies such as AI face-changing. Some people believe that such technologies may have a negative impact on society. In this regard, we contacted Mr. Zhao Lijie, the intellectual property director of Beijing Megvii Technology Co., Ltd. Megvii Technology has rich experience in facial recognition. Mr. Zhao expressed his views on this issue based on his own work experience and the current status of the industry:

Beijing Megvii Technology Co., Ltd.

Intellectual Property Director Zhao Lijie

On the one hand, technology itself is neutral. What we need to limit is not the "AI face-changing" technology, but the improper use of the technology. When ordinary people use "AI face-changing" technology for entertainment and spoofs, they should pay special attention to avoid infringing on others' portrait rights, copyrights and other legal rights, and avoid touching the red line of the law.

On the other hand, the public does not need to worry too much about the phenomenon of some people taking risks and using "AI face-changing" technology to engage in illegal and criminal activities, because there are actually many technical means such as liveness detection technology that can be used to prevent financial fraud and other illegal and criminal activities using "AI face-changing" technology.

Technological development brings new challenges to legal rules

As technology develops rapidly, it also brings new challenges to legal rules.

On April 20 this year, the second draft of the "Civil Code Personal Rights Code (Draft)" was submitted to the 10th meeting of the Standing Committee of the 13th National People's Congress for deliberation. It regulates issues such as human gene embryo research activities, "AI face-changing", human experiments, and personal information protection. The legislative process reflects strong practical significance.

It is understood that during the review and solicitation of opinions on the personal rights section of the Civil Code, some departments have proposed that using information technology to "deep fake" the portraits and voices of others not only infringes upon the personal rights and interests of natural persons, but in serious cases may also have adverse social impacts and endanger national security and social public interests. It is recommended that the law respond to problems such as "face changing" brought about by deep fake technology.

According to the second draft of the Civil Code's Personal Rights Section submitted to the Standing Committee of the National People's Congress for deliberation on April 20, this suggestion was adopted, and a provision was added to the first paragraph of Article 799 of the draft: No organization or individual may infringe upon the portrait rights of others by forging works using information technology.

It is worth mentioning that the draft also stipulates that the permitted use of other personal rights and the protection of natural persons’ voices shall refer to the relevant provisions of this chapter. This means that the protection of natural persons’ voices is also included in the protection of portrait rights.