It has become difficult for Internet users to distinguish between faces created with artificial intelligence technology and the faces of real humans
Baek Kyung-hoon, team manager for AI Yoon produces a video using an AI-generated image of Yoon Seok-yeol, the presidential candidate for the opposition People's Power Party in South Korea. The screen is the level of perfection that has been achieved in fake videos known as "deepfakes" over the past few years.
When the Tik Tok application appeared in 2021 and videos published on it began showing images of American actor Tom Cruise performing a magic trick that makes a coin disappear while enjoying eating lollipops, the name of the account holder was the only thing that indicated to viewers that these clips were not real, and it was The creator of the “Deep Tom Cruise” (or “Deep Tom Cruise”) account on the social networking platform TikTok uses “deep fake” technology to create an AI-generated version of the famous actor performing a magic trick or performing a solo dance.
In the past, the “strange valley” effect was one of the signs of “deep fake” technology, which is the discomfort felt by the Internet user as a result of the hollow look he sees in the eyes of the image created by artificial intelligence, but now we are witnessing the emergence of more realistic images It draws viewers out of the 'Strange Valley' sphere of influence and opens up to them the world of deception all around us thanks to the 'deepfake' technology.
The incredible success of these elaborate images may encourage the use of this technology for criminal purposes, such as the possibility of it being used as a weapon in disinformation campaigns for political or other gain, or the creation of fake porn movies for extortion, or the creation of an endless number of sophisticated tricks intended to harm others. Efforts to devise mechanisms to help us detect fake clips have turned into an “arms race” between digital security professionals on the one hand, and cybercrime practitioners and cyberwarfare activists on the other.
A new study, published in the Proceedings of the National Academy of Sciences USA, shows how advanced this technology is, suggesting that real humans can easily be fooled by artificially created faces—and even more so. “We found that the artificially created faces were very realistic and convincing, and users even considered them more credible than real faces,” says Hani Farid, one of the study’s authors and a professor at the University of California, Berkeley, in the United States. Lots of concerns that "these well-crafted images may be used for criminal purposes."
“We are already facing the dangers of 'deep fakes'," says Piotr Diedek, an associate professor at the University of Switzerland in Italy in Lugano, who was not involved in the study. The tools the study used to create images are readily available to everyone, and while creating a video to the same degree of perfecting the images presents a greater challenge, it is likely that the tools will soon become accessible to everyone, Diedek believes.
The prosthetic faces used in this study were produced through interactions between two neural networks, a type of network known as generative adversarial networks. One research draft after another until it reached the best result. The other network, called the “discriminative network,” which was fed with human images, took over the task of evaluating the artificial faces after comparing them with the data on human faces.
The generative network started using random pixels, and then used the feedback it obtained from the discriminative network to create realistic faces that were successively getting closer and closer to human faces. At the end of this process, the discriminative network was no longer Able to distinguish between real and fake faces.
The networks used in the study were fed a variety of human images including black, East Asian and white faces of men and women, as opposed to the more common procedure in previous research of using images of white men.
After creating 400 images of human faces that matched 400 images of artificially created faces, the researchers asked 315 people to try to distinguish between the real and the fake from among the 128 selected images. On the other hand, the researchers trained another group of 219 participants and gave them some advice about How to spot fake photos while trying to distinguish faces Finally, a group of 223 participants rated the reliability of 128 selected photos on a scale ranging from one (not credible) to seven (completely credible).
The attempts of the first group were random; The members of this group were able to distinguish between real and fake faces with a success rate of 84.2 percent, and the results of the second group did not come much better, as the success rate of participants in distinguishing between real and fake faces was about 59 percent, even with these participants receiving some training and advice to help them in the process selection, and the group that evaluated the authenticity of the images gave the synthetic faces an average rating of 4.82, which was slightly higher than the average rating of the authenticity of human faces of 4.48.
The authors of the study were not expecting these results, says one of the study participants, Sophie Nightingale, "We initially thought that the artificial faces would be less credible than the real ones."
However, it is not possible to say for sure that the “strange valley” hypothesis has been completely retracted, as study participants were able to identify some of the fake images with overwhelming success. The study showed that a significant number of them were unable to distinguish.
The results of the study compound existing concerns about the accessibility of technology that makes almost anyone capable of creating deceptive images. “Anyone can create artificial content without having specialized knowledge of Photoshop or CGI,” Nightingale says. Another problem is that these findings may create the impression that "deepfakes" will soon become irrelevant, adds Wael Abdel Meguid, founding director of the Laboratory for Visual Intelligence and Multimedia Analytics at the University of Southern California, who was not among the study participants. Detectable, and Abdul-Majid is concerned that scientists will refrain from attempts to develop mechanisms that help detect “deep counterfeiting” operations, although he sees the need to develop our ability to detect counterfeiting to keep pace with counterfeiting techniques, and considers that “this is simply another task added to the The role of science in detecting crimes.
“Researchers in this field don't talk much about the importance of working proactively to improve counterfeit detection tools,” says Sam Gregory, director of software strategy and innovation at WITNESS, a human rights organization partly involved in deepfake detection. To the importance of developing fraud detection tools; Because people often tend to overestimate their ability to detect fakes, adding: "The public should be vigilant for the criminal uses of these images."
Gregory - who was not among the study participants - notes that the study authors addressed these issues head-on, and suggested three solutions, including putting static watermarks on artificially generated images, "such as putting fingerprints inside the images so that users know they are artificial images." As he says.
The authors conclude the study with a shocking suggestion. After confirming that the criminal uses of “deep fakes” techniques will continue to pose threats in the future, they say: “Therefore, we urge those who produce these technologies to assess whether the associated harms outweigh the benefits,” adding that “if the harms outweigh the benefits.” We don't like to produce technology just because we can produce it."