Last week in China, Avatarify’s explosion and delisting once again brought AI face-changing technology to the forefront. In fact, this technology has already appeared, and it is not the first time that people have been so crazy about AI face-changing related applications.
For example, the most popular selfie modes a while ago-"Disney princess face" and "two-dimensional comic face" were generated using AI.
Last year, Tencent's light and shadow research room launched various AI special effects such as "king face" and "fairy tale face" on social platforms such as QQ Small World, Weishi App, and Weibo. It combines the generation of confrontation network GAN algorithm with 3D cartoon style and game character style, which can immediately turn the characters in photos and live videos into game characters, real "Barbie" or "Princess on the Run".
Google's AI platform is also based on a generative confrontation network, combining face detection, hair segmentation, portrait segmentation and other technologies to launch portrait animation services. It has now cooperated with B612 Kaji Beauty Camera to provide users with customized two-dimensional comic avatar customization. Many anchors and Internet celebrities have used this kind of special effects in live broadcasts, using this kind of stickers for self-shooting, and it also forms a trend as a new "coding" method.
AI face change heat:
The rise of AI face changing can be traced back to 2016. The German research team developed an application called Face2Face that can realize real-time facial image conversion. In 2017, a user named Deepfake used AI face-swapping technology to "graft" the face of "Supergirl" Gal Gadot to the role of an adult movie, causing an uproar. After the code was open sourced, this technology, collectively called DeepFake, became popular.
The domestic "AI face-changing fever" began with a video on Bilibili in 2019 that replaced Huang Rong played by Zhu Yin in the 94 edition of "The Legend of the Condor Heroes" with Yang Mi. At that time, the Weibo topic # replaced Zhu Yin's Huang Rong with Yang Mi's face# has read over 100 million. Although this film was taken off the shelves due to infringement of portrait rights, it could not stop the craze of AI changing faces.
At the end of August 2019, Momo launched ZAO, which focuses on AI face-changing video production. This application allows users to "transplant" their faces to the celebrity demo, and quickly became the first place in the Apple App Store free application list after it went online.
However, the good times are not very common. Less than a week after the launch, ZAO's operators have raised doubts from all walks of life due to the irregular user privacy agreement and the existence of data leakage risks and other network data security issues. The application was also interviewed by the Ministry of Industry and Information Technology, and the application was also taken down.
Last week, a short video with "Ma-ia hi~Ma-ia hu~Ma-ia ho~Ma-ia ha-ha..." this cheerful, brainwashed melody as the background music was swept wildly on major social media. Screen.
In popular online films, Musk, Zuckerberg, Buffett, Jack Ma, Ma Huateng and other business tycoons make exaggerated expressions and shake their heads; Eason Chan, Donnie Yen, Wu Jing, Charlene Choi, tranquility and other celebrities; even include Mongolian Characters and small animals in the paintings such as Na Lisa, cats, dogs and dogs.
This magical special effect was produced, and the best-selling old song "Dragostea Din Tei" released in Romania in 2003 was once again spread around the world. It is an AI face-changing software called Avatarify.
Avatarify was originally a joke developed by Russian programmer Ali Aliev to tease his colleagues who had video conferences together. This application can "graft" other people's photos onto one's own face, and can also manipulate the expressions of people in the photos by inserting still faces into video demos to make them "moving" and generate short videos.
Last July, Avatarify officially appeared on the Apple App Store. In mid-February this year, the ranking of this app in China suddenly rose, and on February 25th, it surpassed WeChat, Weibo, and Douyin, becoming the first in the free app download list of the IOS App Store for many days. "Domination".
After seven days of popularity, Avatarify has now been removed from the App Store in China First, and the official has not explained the reasons for the removal.
Even if ZAO and Avatarify are short-lived, and DeepFake is restricted by the US Congress, human enthusiasm for "face-changing" technology continues to rise. After all, the "viral" spreading effect caused by such topic-filled videos is an excellent channel for the flow-dependent film and television creation, e-commerce shopping, and short video social industries. In recent years, AI's transformation of "face" is still mainly used in social entertainment scenes. With the development of related technologies, various face-changing and face-making apps and services are also emerging in endlessly.
According to the observation of "Intelligent Relativity", the existing AI face-changing related applications can be roughly divided into three categories: The first category is the "virtualization of human faces that are more mature for commercial applications such as Tencent's "fairy tale face" and google"portrait cartoonization" ", that is, through face recognition and image generation technology, the user's face becomes a customized cartoon image.
Since the "cartoonized" image is highly entertaining, it is quite different from real people and is easy to distinguish, so this type of technology is less controversial. Major manufacturers sell such technical services to the B-end, forming a relatively mature business model that pays for the amount of use or different functions.
The second category is like Avatarify, which makes "almost true" expressions and face replacements on the faces of human figures. It is generally charged through the free value-added service model, or as a means of attracting customers to guide users to pay for other functions.
For example, Deep Nostalgia (Deep Nostalgia) developed by the online genealogy company MyHeritage is an AI application with the function of "generating dynamic expressions". This service can make old static photos into small videos of a few seconds, and use technology to allow users to watch the lifelike expressions of the deceased, hoping to comfort users' longing for their loved ones.
If you want to make a still photo move, you need to use the generative confrontation network GAN to denoise and reconstruct the low-frequency image, complement the content of the high-frequency details in the image, and increase the resolution of the image to make it clear. The current photo restoration technology also uses this kind of depth generation algorithm to restore image details, even the mosaic part can be restored.
Next, use the motion migration algorithm to extract the key points in the video characters in real time using the trained AI model, and mark the corresponding key points for the faces in the photos, and then correspond the two to make another segment. The actions of the characters in the film are transferred to the characters in the photos.
Now, MyHeritage's AI module can be used on Twitter, as a marketing tool to cause Twitter users to follow the trend and nostalgia. Tencent Microvision also launched the "Old photos can be moved" function. By uploading photos, you can create a dynamic image based on the content of the photo, just like the magical photos in Harry Potter.
There have also been videos and tutorials on Bilibili and Youtube that use AI technology to make the portraits in a series of paintings such as Du Fu, Zhu Yuanzhang, Xu Beihong, Lin Daiyu, the girl with pearl earrings, Mona Lisa, and Venus "live". There are also up owners in China who use google Feida's open source projects PaddleGAN and DFDNet to "restore color" and make dynamic facial expressions in black and white photos of people.
DeepFake that can overwhelm the boat:
Whether it is Avatarify, Deep Nostalgia or ZAO, the controversial reasons for these AI face-changing applications are due to the leakage and abuse of personal privacy caused by the fake and authentic DeepFake technology and facial recognition.
Since the emergence of Generative Adversarial Networks (GAN) in 2014, computer image and film synthesis technology has continuously achieved amazing results. Although the technology is neutral, since 2017, when Reddit users used this technology to fake pornographic videos of Supergirl Supergirl, Deepfake fraud gradually penetrated into politics, entertainment, media and other fields, and even threatened the US presidential election.
According to a survey by security analysis company Sensity, since December 2018, the number of Deepfake counterfeit products has increased year by year, and has doubled in about half a year. As of June 2020, there were 49,081 fake films, a year-on-year increase of more than 330%. Among them, pornographic films accounted for the highest proportion, and most of the target objects came from the film and television entertainment industry, politics and business.
Last year, a Malaysian politician was sentenced to jail for a film suspected of engaging in homosexual activities, but it was later confirmed that the film was a Deepfake fake. Although funny videos produced on the basis of this technology have brought some joy, as far as the results of current use are concerned, the impact of Deepfake technology on people seems to be more harmful than good.
Therefore, Internet giants such as Facebook, Google, and Microsoft have begun to devote themselves to developing detection technologies that can compete with Deepfake. Facebook released the largest Deepfake database ever, and even held a Deepfake Detection Challenge (Deepfake Detection Challenge). The third category of applications related to AI face changing-"Fake Face Screening" products and APIs were born.
Since the naked eye can hardly distinguish the more mature Deepfake fake films, it is necessary to rely on AI to crack the AI face-changing technology. Existing methods such as detecting the heartbeat, blood flow and other biological signals of the video target, by analyzing the residual of the signal to determine the true and false, the current accuracy of this method has reached 97.29%, and it can also detect the deepfake behind it. model.
Domestic major manufacturers have also launched such applications, such as the "face detection", "H5 video live detection", and "online live detection" APIs launched by google's brain AI open platform , which have the ability to identify facial composite images. Based on deep learning algorithms, it can identify image features such as moiré and imaging malformations, prevent cheating attacks such as static pictures, screen or video remakes, and improve business security. At present, this function has been widely used in real estate transaction identity authentication; insurance insurance, underwriting and claims auxiliary verification; real-name card issuance by operators, car leasing and other businesses.
Alibaba Security Turing Lab has also created a detection algorithm Sharp-MIL (S-MIL) based on multi-instance learning for this purpose. Only video-level annotations are required to detect deepfake works.
Tencent Cloud AI Vision launched AntiFakes fake face recognition technology. Based on image algorithms and visual AI, it can detect and analyze the authenticity of faces in pictures or videos, and identify whether the faces in the pictures are generated by AI face-changing algorithms and apps. The fake face, and finally the risk level of the image or video is evaluated.
Emotional Identification Network (Emotion Recognition Network) is one way of measuring whether the scene context or audio content matches to determine the authenticity of the film through the detection face emotions. There are also background difference detection that compares the face area and the background area, looking for small differences to distinguish between true and false; and remixing the fake face into the target image or movie to find the boundary from the face X-ray detection (Face X-ray) . After testing on the FaceForensics dataset, the accuracy of face X-ray can reach more than 99%.
Although the AI face-changing technology based on Deepfake can generate huge traffic, it can be used as a good marketing tool. Before the development of a corresponding regulatory system, such applications that cannot ensure the security of user privacy data still cannot escape the fate of being removed, but instead bring many business opportunities for Deepfake prevention software. In addition to the above products, there are commercial application programming interfaces such as Sensity, Deepware, and Microsoft Video Authenticator.
AI face-changing technology can be said to open Pandora's box. While bringing entertainment to some people, it also becomes a tool for criminals to harm others. How to properly use this double-sided blade to strike a balance between entertainment and privacy is a problem that mankind must face in the process of technological development.