The sci-fi movie "Blade Runner" depicts a future world that is difficult to distinguish between true and false:
▲ Who are the replicators? Image from: "Blade Runner 2049"
The "replicants" made by humans have almost identical faces to humans. Unless a professional instrument is used for emotional testing, or the number hidden in the body is found, humans cannot identify "replicants" from their appearance at all.
Such a future world may be a little far away, but if you only look at the face, you may really not be able to tell the truth from the fake.
AI "fake face" has crossed the uncanny valley
First look at a set of pictures, do you have the confidence to distinguish the "fake face" at a glance?
Which of the above faces is fake? (See the end of the article for the answer, but don’t worry, there are still many faces to come)
The concept of "fake face" mentioned here refers to face photos synthesized by AI, not real people. If you feel that it is difficult to judge true and false for a while, don't worry, you are not alone.
Professor Hany Farid of the University of California, Berkeley, has been working on AI image synthesis technology for many years.
According to a study he recently published in the Proceedings of the National Academy of Sciences, AI-synthesized faces are indistinguishable from real people, and even look more trustworthy than real people.
The results of this study were beyond the researchers' expectations.
Dr. Sophie Nightingale, who co-authored the research, said that the initial purpose of the research was to find ways to improve the credibility of AI faces using real-life comparisons.
Farid believes that the current rate of development and improvement of AI image synthesis technology is very rapid, even faster than traditional CG imaging.
We think we've crossed the uncanny valley of static faces
The Uncanny Valley Effect is a psychological theory that people will have a favorable impression of objects similar to people, but when the similarity reaches a certain level (such as zombies, anthropomorphic dolls), people's reactions will become extremely negative and disgusting.
When the similarity further rises and reaches the similarity with the real person, the emotional response of people will return to positive, and there may be an empathy effect between humans.
Judging from the experimental results of Farid, the face synthesized by AI is likely to have left the stage of "Walking Dead".
How is such a realistic human face synthesized?
Generative Adversarial Networks (GANs) are currently the most mainstream synthesis algorithms. The name sounds very unfamiliar, but its logic is not complicated.
To put it simply, there is a "painter" and an "appraiser" in the GAN. The "painter" needs to draw a picture that resembles a human face as much as possible, and give it to the "appraiser" to judge.
The "appraiser" looks at a lot of real photos before judging and analyzes the characteristics of the face. When the "painter"'s painting can fool countless "appraisers" who have read faces, an AI synthetic face photo is born .
In this process, the accuracy of the "appraiser" will continue to improve after continuous learning, and the corresponding skills of the "painter" will also increase. The two form a confrontational relationship, thereby improving the quality of the synthesized image until it can match the fake to the real. .
Farid used the Nivdia StyleGAN2 model released by NVIDIA in the experiment. In order to study the credibility of its synthetic photos, the researchers organized three experiments.
In the first experiment, the researchers invited 315 participants to distinguish 128 groups (a total of 800 groups) of photo groups composed of real people and AI. As a result, the average correct rate of participants was less than 50%, only 48.2%.
Then in the second experiment, the researchers invited 219 new trained participants to participate in the same experiment and gave correct feedback after each parting.
With the help of the researchers, the accuracy of the second experiment was improved, but it was only slightly over 50%, reaching 59.0%. Farid and Nightingale were not surprised by the fidelity of the AI photo, but the result of Experiment 3 was unexpected.
In the third experiment, 223 new participants rated the reliability of the same batch of photos from 1-7. The results showed that the reliability of AI-synthesized photos was 7.7% higher than that of real faces. This tiny The gap is statistically significant.
▲ Deepfake face-changing Nicolas Cage
The researchers believe that AI-synthesized photos that have a certain degree of credibility are likely to be used by criminals in the future to commit fraud or cause confusion in social networks. This issue needs to be paid attention to by the society, and the development of image synthesis technology requires some agreements. constraint.
So the question is, since AI synthetic face has certain social risks, why do people still invest energy in research?
AI face is very good, but it is a "double-edged sword"
▲ Keanu Reeves debuts
At the 2019 E3 video game show, Keanu Reeves' surprise appearance in the "Cyberpunk 2077" trailer instantly ignited the emotions of all audiences inside and outside the venue, and then he himself came on stage and announced that he would play an important role in the game "Johnny Silver" "Hand" face model, aroused the expectations of countless players.
▲ In Resident Evil 8, Helena Mankowska served as the face model of Mrs. Timitrescu
Because in the virtual world, a realistic human face can bring a strong sense of immersion to players. With the improvement of game functions, using real people as face models instead of digital pinch faces has become more and more game manufacturers to shape characters. means.
But using a real face model often means high portrait licensing fees and motion capture costs, which are not realistic for some small studios.
This is where a copyright-free AI synthetic face comes in handy—it sounds plausible that a person who doesn’t exist in the world plays a virtual character.
For example, Generated Photo, a free AI composite photo project, has cooperated with animation software company Reallusion to use AI-synthesized portraits as 3D images for animation, games or advertisements. Developers can freely choose the race, age, and gender of the characters. There are copyright issues.
Just imagine, if the faces of NPCs in games such as Sims or GTA are made of realistic portraits, the immersion and presence of the game will be greatly improved.
In addition to games, some customer service software also requires a large number of real avatars to communicate with customers. If you replace real avatars with AI avatars, you can not only avoid portrait copyright disputes, but also protect personal privacy from being leaked.
However, although AI-synthesized photos have reasonable significance, at the same time, they will also have a certain impact on the authenticity of online photos. After all, no one wants to be fascinated by a non-existent person on a dating app.
Farid believes that the only way to solve this problem is to add a "authenticity" certification to each photo that is actually taken, so that people can distinguish its authenticity when viewing and using the photo.
It sounds like a reversed version of Blade Runner, where people in the movie have a numbered identification carved into the eyeballs of the replicants, while in reality they are tagged in real photos to combat the "fakes".
In fact, companies such as Adobe and Microsoft are already promoting related technologies.
In February 2021, companies and institutions including Adobe, Microsoft, Intel, Arm, and Truepic formed a project called the Content Origination and Authenticity Alliance (C2PA) to combat disinformation and create a verifiable image authenticity technical standards for safety and traceability.
The verification method is also very straightforward – the information such as the shooting and post-modification of the photos are kept intact with blockchain technology, and no matter how the photos are modified, they can be viewed directly.
Certified real photos will have a small "i" in the upper right corner. When you click it, you can see the details of the shooting date, location, and lens generated by the camera. If someone uses software such as Photoshop After modifying the photo, you can also revert to the original image of the photo to get the full image.
The C2PA certification method can ensure the authenticity of photos in serious news and other fields to a certain extent, but due to its short establishment, it is currently only used on some media or social platforms, and it is still outdated to provide authenticity guarantees for all Internet content morning.
That is to say, in the future, AI synthetic photos may be a social security risk. Currently, image synthesis models such as Nivdia StyleGAN2 can be publicly downloaded on platforms such as Github. Is it really safe to do so? Farid believes that this requires careful consideration by technical personnel after balancing the benefits and risks.
So the question is, there are so many faces in the article, which ones are real people?
The answer is: Unless otherwise marked, all are false.
#Welcome to pay attention to the official WeChat account of Aifaner: Aifaner (WeChat: ifanr), more exciting content will be brought to you as soon as possible.