In September, a doctored video of the National Journal, the main news on the Globo television network, won social networks. In it, presenters William Bonner and Renata Vasconcellos showed the results of a poll of intention to vote for the Presidency, but the data were inverted about who was the favorite candidate, both in the graphics and in the presenters’ speeches. The following day, the newscast made a clarification warning that the video was being used to misinform the population and stating that it was deepfake, a technique that uses artificial intelligence to make deep edits to the content. With it, it is possible, for example, to digitally change a person’s face or simulate their voice, making them do what they didn’t do or say what they didn’t say.
In August, another newscast video with a similar edition, which also inverted the results of a poll for the Presidency, was posted on the social video network TikTok, where it reached 2.5 million views, according to Projeto Comprova, an initiative that brings together journalists. of 43 media outlets in the country to check misinformation.
“Some deepfake technique may have been used in these videos, but a more detailed analysis is needed. For us, the important thing is to know that they are fake”, observes computer scientist Anderson Rocha, director of the Institute of Computing at the State University of Campinas (Unicamp), where he coordinates the Artificial Intelligence Laboratory (Recod.ai). The researcher has been studying ways to detect malicious tampering in photos and videos, including in deepfakes, also called synthetic media.
In March of this year, shortly after the start of the war between Russia and Ukraine, Ukrainian President Volodymyr Zelensky was a victim of deepfake. A video in which he appeared to urge Ukrainians to lay down their weapons and return to their homes, as if the country was surrendering, circulated on social media, forcing Facebook and YouTube to remove them as soon as they were found to be fake. In the images, the president’s face appeared in a body that hardly moved, dressed in a green T-shirt.
In some cases, as in the videos of the National Journal, it’s not that hard to see that they’ve been altered in some way, because the original news is easily available for verification. But this is not always the case. In the face of synthetic media, the saying “seeing is believing” is losing its meaning, and artificial intelligence itself can be an ally.
“Usually synthetic videos are made in two stages: first, using a deepfake platform, to change faces or synchronize the mouth, and then editing is done in video editing programs”, explains Rocha. Who knows what to look for usually detects some flaw in the program used to produce the farce, like a game of different lights, a contrast between the original video and the new face that was inserted.
It’s like cutting a face out of a photo and placing it on top of another: the incidence of light and the way the camera captured the two images are different. These vestiges are clues that lie along the way, identified by computer forensics techniques, a research area that has grown in recent years and of which Rocha is a part.
With colleagues from the University of Hong Kong, the researcher developed an algorithm that helps to detect, simultaneously in the videos, if there was manipulation of faces and, if so, to locate which regions were changed. It could, for example, have been the entire face or just the mouth, eye area or hair. “The average of correct answers was 88% for low resolution videos and 95% for videos with higher resolution”, explains Rocha, about a universe of 112 thousand tested faces: half real, half manipulated and generated by four deepfake programs. The method also indicates whether the image was created from scratch, that is, not based on an existing photograph. The results were published in April 2022 in the journal Transactions on Information Forensics and Security.
According to the computer scientist, other algorithms developed are able to detect traces of alteration in deepfake videos, but, for the most part, they work based on the clues left by better-known manipulation programs, basically divided into two categories: those that allow the exchange of faces and those that allow editing of facial expressions. One such software is known to always leave some imperfection in the mouth synchronization – the detector algorithm is then programmed to look for that specific error. “There’s a problem with that: if we don’t know the deepfake software, it becomes more difficult to identify these traits. And there are always new applications,” says Rocha.
So he and his colleagues trained the algorithm they developed to detect clues without assuming knowledge of the deepfake generator application. “We work with the idea that, regardless of the program, it will leave a noise, something that is not coherent with the rest of the image.” The method works on two fronts: it looks for noise signatures, that is, subtle changes in the edge of the face, for example, and it maps the so-called semantic signature, which can be a color, texture or shape flaw.
“The algorithm automates the process that a human expert does, which is looking for inconsistencies, like contrasts in light,” he says. “The next step is to test it with fake videos generated by a larger number of programs, to confirm this potential.”
Pornography and the rise of fake news
This type of detector algorithm can be used for a variety of purposes involving combating the malicious use of deepfakes. Rocha is part of an international project called Semantic Forensics, alongside researchers from the Universities of Siena and Polytechnic of Milan, in Italy, and of Notre Dame, in the United States, which has the support of the United States Department of Defense. The objective is to develop automated tools that detect these manipulations. “We have already seen cases of altered videos of military exercises in other countries, which multiply the number of missiles to show greater war power”, he says.
These algorithms can also help in cases of political deepfakes, such as the Ukrainian president episode, or even pornographic ones. It was using sex movies that the technique gained fame, at the end of 2017. At the time, some internet users began to insert the faces of movie celebrities in scenes of films with sexual content. In September 2019, according to a survey by the DeepTrace Labs, a Dutch cybersecurity company, 96% of the deepfake videos mapped on the network were non-consensual pornography. The main victims were women, especially actresses, but there were also records of cases with people who were not famous. In July of this year, the singer Anitta was also the victim of a deepfake porn. The original video had already been used to produce fake media with the face of actress Angelina Jolie.
According to journalist Cristina Tardáguila, program director at the International Center for Journalists (ICFJ) and founder of Agência Lupa, which specializes in fact-checking, Brazil has already dealt with deepfakes that need to be debunked. Therefore, programs that help detect synthetic media can be allies of journalists and fact checkers, who work against time. “When dealing with disinformation, you have to be quick. Therefore, it is important to invest in artificial intelligence, in tools that can help detect and map this type of fake content more quickly. Thus, we were able to shorten the time between the propagation of false content and the delivery of the check”, she evaluates.
“Deepfakes are the height of fake news. They have the potential to deceive more easily, because, if it is a video, the person is watching that scene”, observes journalist Magaly Prado, who is doing a postdoctoral internship at the Institute for Advanced Studies at the University of São Paulo (IEA- USP). “Audio can also be generated synthetically,” she says, author of the book Fake news and artificial intelligence: The power of algorithms in the disinformation war released in July by Edições 70.
She assesses that, despite being less remembered and less common, deepfakes exclusively in audio format have the potential to spread through platforms such as WhatsApp, a media widely used by Brazilians. They follow a logic similar to that of videos: with accessible apps that get better and better, it is possible to simulate someone’s voice. The easiest victims are public figures, whose voice is widely available on the web. The technique can also be used for financial scams. “There have been cases like a technology company employee who received a voice message from a top executive asking for a cash transfer. He got suspicious, the case was analyzed by a security company and it was found that it was a message built with artificial intelligence”, he says.
Journalist Bruno Sartori, director of the company faceFactory, explains that producing well-made deepfakes, both audio and video, isn’t that simple – yet. That’s what he does professionally: his company creates synthetic media for commercial use and provides content for comedy shows on TV channels Globo and SBT.
In 2021, he worked on a commercial for Samsung in which the adult presenter Maisa interacted with her child version. The latter, created with deepfake techniques. The virtual little girl dances, plays and throws a notebook in the air. On another occasion, he had to insert an actor’s face into a stunt double. “To train artificial intelligence well, it is important to have a good bank of images and audio of the person you want to imitate. Good programs that do high-quality processing also need to have advanced settings. Otherwise, there could be visible flaws in the face or, in the case of audio, a robotic voice,” he explains.
In his opinion, the manipulated videos of the National Journal that exchange research data were not changed using artificial intelligence. “In my analysis, they went through a common editing, with cutting and reversing the order of the audios. They are what we call shallowfake. But, as it is well made, the potential to deceive people is the same”, evaluates Sartori. For him, in a few years these programs will be lighter, smarter and more accessible.
There are a few ways to protect yourself from disinformation created with the help of technology. One of them is to pay attention to the use and privacy licenses of the most diverse free applications used on a daily basis – from those that ask for access to the user’s photos to generate fun effects, through to those that can store the voice. According to Rocha, from Unicamp, many of them keep a large amount of data that can be shared for other purposes, such as training deepfake programs.
Another important point is media education. “As much as there is software that helps us point out what is false, the first step is to distrust what you receive on social networks. And check the sources of information, research them”, he concludes.
Project
Déjà vu: Temporal, spatial and characterization coherence of heterogeneous data for integrity analysis and interpretation (nº 17/12646-3); Modality Thematic Project; Responsible Researcher Anderson Rocha; Investment BRL 1,912,168.25.
Scientific article
KONG, C. et. al. Detect and locate: Exposing face manipulation by semantic- and noise-level telltales. Transactions on Information Forensics and Security. v. 17. Apr. 2022
Book
PRADO, M. Fake news and artificial intelligence: The power of algorithms in the disinformation war. São Paulo: Editions 70, 2022.
Text: Sarah Schmidt, from Pesquisa Fapesp Magazine