When you see this photo, do you think it's a real human face photo?
In fact, it's a synthetic face photo generated by a website called "This person doesn't exist" (this-person-does-not-exist.com).
"Our assessment of the realism of AI synthetic face photos shows that the compositing engine has gone through the 'uncanny valley' and is able to create faces that are indistinguishable from real faces and are more trusted." On Feb. 14, a paper published in the Proceedings of the National Academy of Sciences (PNAS) stated in the abstract.
The "uncanny valley effect," proposed by Masahiro Mori in 1970, is a hypothesis about how humans perceive robots and non-human objects.
The "uncanny valley effect" said that because robots and humans are similar in appearance and movement, humans will also have positive emotions for robots; and when the similarity between robots and humans reaches a certain degree, human reactions to them will suddenly become extremely negative and disgusting, even if robots and humans are only a little different, they will appear very conspicuous, so that the whole robot is very stiff and terrifying. In scenes where composite face photographs are synthesized, the "uncanny valley" effect often comes from the uneasiness caused by the hollow expression in the synthetic person's eyes.
Once the similarity between robots and humans continues to rise, equivalent to the similarity between ordinary people, the emotional response of humans to them will return to the positive, resulting in empathy between humans and humans.
Increasingly convincing images are pulling viewers out of the "uncanny valley" into a world of deception constructed by Deepfake (Deep Forgery). In the study "AI synthetic faces are no different and more trusted" by Hany Farid, a professor at the University of California, Berkeley, and Sophie Nightingale, a doctoral student at Lancaster University, the participants in the experiment were asked to distinguish between the neural network StyleGAN2 synthetic faces and real faces, and the level of trust that these faces evoke.
The study consisted of three experiments. In the first experiment, 315 participants were classified as real or synthetic faces from 128 faces (extracted from a group of 800 faces) with an accuracy rate of 48 percent.
In the second experiment, 219 new participants were trained in how to recognize real faces versus synthetic faces, and then classified 128 faces as in the first experiment. Despite the training, the accuracy rate was only improved to 59%.
The researchers then decided whether exploring the perception of trustworthiness could help people identify artificial images, "Faces provide a rich source of information that takes just a few milliseconds to make implicit inferences about personal characteristics, such as trustworthiness." We wondered if synthetic faces would activate the same credibility judgment, and if not, then the perception of believability might help distinguish between real faces and synthetic faces. ”
In a third experiment, 223 participants scored the credibility of 128 faces taken from the same group of 800 faces ranging from 1 (very implausible) to 7 (very trustworthy). Finally, the average score of the synthetic face is 7.7% higher than the average score of the real face, which is statistically significant.
The entire experimental results show that the synthetic face photo is almost indistinguishable from the real face, and is even considered more trustworthy. Such results were also unexpected by the researchers, Nightingale said, "We initially thought that synthetic faces were not as credible as real faces." ”
This StyleGAN, which generates face photos, is a neural network developed by Nvidia in 2018. GANs consist of 2 competing neural networks, one called a generator, constantly generating something and the other called a discriminator, constantly trying to determine whether the result is real or generated by the first one. The generator starts the exercise with random pixels. With the feedback from the discriminator, it gradually produces an increasingly realistic human face. Eventually, the discriminator was unable to distinguish between a real face and a fake face, and the training was over.
Creating non-existent face photos is actually a by-product of GAN, whose original main goal was to train artificial intelligence to recognize fake faces and general faces, and Nvidia needed to improve its graphics performance by automatically recognizing faces and applying other rendering algorithms to them. However, because the StyleGAN code is public, one of Uber's engineers used it to create a random face generator.
The malicious use of Deepfake (deep forgery) technology has been reflected in reality, such as the false propaganda campaign in the US election, the fake pornography created for extortion, and so on. Since the advent of Deepfake technology, the identification of deep forgery and further deceptive identification has become an "arms race."
Now the study on Deepfake's progress raises concerns about its misuse, "anyone can create synthetic content without Photoshop or CGI expertise," Nightingale said.
Wael Abd-Almageed, director of the Visual Intelligence and Multimedia Analysis Laboratory at the University of Southern California, said in an interview with Scientific American, "Another concern is that these findings will make deep forgery completely undetectable and that scientists may abandon attempts to develop countermeasures against deep forgery." ”
The two researchers also proposed countermeasures, such as merging powerful watermarks into image and video synthesis networks, which will provide an effective mechanism for reliable identification.
Perhaps most pernicious, the paper writes, is that in any digital world where images and videos can be forged, the authenticity of any unpopular record can be called into question. "So we encourage people who are advancing technology to consider whether the risks outweigh the benefits, not just whether they are possible from a technical perspective."