Contents
We have talked about it a lot, Mark Zuckerberg recently paid the price for a Deepfake. This tendency to falsify a video using AI. Until now, this phenomenon only used videos, or audio recordings. Now, software is able to animate a person’s face from a single photo. All while making him say what we want. This new tool has not finished fueling the current frenzy around Deepfakes.
This AI is able to bring famous portraits like the Mona Lisa to life, or Marilyn Monroe. Their eyes will be able to blink again, but above all their lips will come alive to pronounce coherent words, or not. Researchers from the Samsung AI Center, and the University of Imperial College London, are behind this project. They have already enabled the creation of a few extracts with not yet very conclusive results. For example, Einstein who evokes the wonders of science is not very credible, well always more than Rasputin who plays Beyoncé’s Halo.
Microsoft acquires Nuance, the specialist in speech recognition, for $ 19.7 billion
This Deepfake uses an audio recording of Einstein, then one of his photos. These two elements mixed by the AI make this result.
The second example is more funny since the researchers used Beyoncé’s voice, on the portrait of Grigory Efimovich, more commonly known as Rasputin.
Aside from these two examples that are not credible enough not to take them seriously, this technology can be more realistic and present real threats in terms of fake news. In the video below, people are saying trite phrases like “it’s 11am” or “don’t forget your jacket,” with a tone, and facial expression completely in disagreement. This is unnatural, but well handled by AI.
While not all convincing, these videos are impressive from a technical standpoint. The speed with which work in AI is progressing should allow the creation of ever more credible content. Imagine these Deepfakes in case of election campaigns, the results could be disastrous. Donald Trump has already started to disclose these fake recordings, as seen with the video of House Speaker Nancy Pelosi.
Researchers create AI that makes it easy to create deepfakes from a single image
Currently being researched to make them more efficient, deepfakes – a contraction of deep learning and fake – allude to image synthesis techniques based on artificial intelligence. If they are recent enough, current deepfake techniques are already giving very impressive results, which worries US lawmakers.
There are concerns that they may be used against people to represent them in situations they would oppose, such as in ___ographic videos. It should be noted that this was the original purpose of this technology: it can be used in industry p *** o, where faces can be assumed to be others. According to US lawmakers, deepfakes also represent a real political danger. They think, for example, that this image synthesis technique could be used as part of wider disinformation campaigns with the aim of influencing elections by disseminating fake news which could then appear more credible. In other words, deepfake creation technologies could be dangerous weapons in the hands of malicious people.
Despite fears that deepfakes may be used for malicious purposes, research to make this technology more powerful is increasing. The latest is the result of the work of a team made up of researchers from a Samsung artificial intelligence center and the Skolkovo Institute of Science and Technology, both based in Moscow.
Researchers have devised a way to create deepfakes or “living portraits” from a very small set of data – a single photograph is sufficient in some cases. Their research titled “Few-Shot Adversarial Learning of Realistic Neural Talking Head Models” was published Monday in arXiv, the archive of electronic preprints of scientific papers.
They use machine learning techniques called few-shot or one-shot learning, which allude to learning from a few examples or from a single example. Indeed, if artificial intelligence is efficient in applications using a lot of data, it lacks the ability to learn from a limited number of examples. In other words, current deep learning methods suffer from low efficiency for small samples, which is in stark contrast to human perception: even a child can recognize a giraffe, for example, after viewing a single photo. The few / one-shot learning therefore aims to endow deep neural networks with this ability: to achieve a high level of learning from very small data sets.
In their work, Samsung researchers were therefore able to use an image to create a compelling animated portrait. By increasing the number of shots (passing to 8 or 32 photos for example), the animated portrait becomes more and more realistic. And it even works on Mona Lisa’s portrait and other unique photos. In a video posted by one of the researchers on YouTube, the famous portraits of Albert Einstein, Fyodor Dostoyevsky and Marilyn Monroe come to life as if they were videos that had been captured live from your camera. smartphone.
Just because Samsung’s AI needs a single photo to generate a living or deepfake portrait doesn’t mean the algorithm relies on just that single photo. Behind, there is some preliminary meta-learning work that has been done. The researchers have indeed carried out a long training on a large database of a public repository of 7,000 images of celebrities extracted from YouTube. This allows the algorithm to identify what they call “historical” features of faces: eyes, mouth shapes, length and shape of a nasal bridge.
From there, the system is able to generate a video or an animated portrait from photos of people who were not taken into account in the meta-learning phase. Concretely, when the system receives the photo of a new person as input, it initializes its parameters in a way specific to the person, so that the generation of the deepfake or animated portrait can be based on a few images and be done quickly, “Despite the need to adjust tens of millions of parameters”.
As with most deepfakes, there are a few flaws. Most of the faces are surrounded by visual artifacts. But the results remain impressive given that this technique is in its infancy.
This research represents a significant advance over what other deepfakes and algorithms using generative adversarial networks can accomplish. Instead of teaching the algorithm how to paste one face onto another using a catalog of one person’s expressions, the researchers use the facial features that are common to most humans to then puppet a new face.
“Several recent works have shown how extremely realistic human head images can be obtained by forming convolutional neural networks to generate them. In order to create a custom model of the head of a person who is speaking, this work requires training on a large dataset of images of the same person. However, in many practical scenarios, such models can be formed from a few images of a person, or even a single image, ”the researchers explain. “We show that such an approach makes it possible to create very realistic and personalized models of new characters and even portraits. “
Now anyone can create deepfakes of a quality beyond belief
Deepfakes are previously digitally edited videos. They allow, thanks to an AI, to place a face on another face. At the heart of the news, they impress and entertain. Nonetheless, this technology can have a detrimental influence and is the cause of much concern, especially since anyone can now create deepfakes of a quality beyond belief.
In the video above, Dr. Károly Zsolnai-Fehér’s explains that anyone can now create deepfakes. In order to use this technique, he explains that you just need to film yourself and add the face of your choice. In a deepfake, several movements are transcribed, such as those of the head, the mouth or the eyes. But, it is also possible to move the whole body.
In addition, there are many techniques today for creating deepfakes. One of them consists, for example, in adding several references of the face or a particular pose of a video of the chosen person. The latter does not require any special knowledge of image editing. This technique is therefore very general. It is also possible to create a high quality deepfake from a single photo of the selected face, or have us dance like professionals. You can also do this with images of animals and characters or even realize an animation of a robot from a single previously filmed sequence.
It is a neural network that makes it possible to generate all this information. This one manages to identify what kinds of movements and transformations were made in a video. He then manages to automatically create a deepfake.
“I think the first step is to inform the public that these deepfakes can now be created quickly and with little means and that they no longer need a graduate scientist. If this can be done, it is of the utmost importance that we all know it ”, explains Károly Zsolnai-Fehér’s in particular.
