Francesco Angeli wakes up every morning at six to do yoga. He meditates in Bali, in one of the most fascinating places on Earth. But he didn’t end up in Indonesia looking for himself. Francis, engineering graduate, he recounts the “too many times” that he has heard from potential customers: “If we send this thing to Korea or China, they can do it in less time and spend less”. “So you know what I did? – he adds – I went to Southeast Asia too ”.
Ironically, Angeli flew about 12,000 km away to continue working in Italy. Antonio Riccithe creator of Striscia la Notizia, has entrusted him a complicated challenge: “Going beyond the limits of real imitators – says Angeli – creating perfect copies of politicians or celebrities through digital”. The Italian engineer succeeded in making the most of it the potential of deepfakewhich is basically a digital portrait: “It is a sequence of moving images in which the subject is portrayed by a software” explains Angeli. In a movie, for example, one person’s face can be replaced with another’sto your liking.
Thus was born the ‘masterpiece’ of Striscia: the digital copy by Ilary Blasi. So perfect – compared to other deepfakes in circulation – that at first glance one would lead one to believe that that recovery be the original. Not so, of course: behind the digital copy is the face of the imitator Francesca Manzini. And Francesco Angeli explained to us how it was possible to superimpose Blasi’s face so accurately.
Why is Ilary Blasi’s deepfake so special?
“Ilary is an exceptional case that I would call a metafake, because it is a fake that speaks of itself. Ilary has such a precise face, so symmetrical, so harmonious. He has an almost non-existent skin texture – clearly due to makeup – which results in a lower than average level of information in his face of the information and details present in the faces of many other characters portrayed in the past by Striscia. So, in reality, artificial intelligence has worked a lot less on Ilary because. The real difficulty was convincing the AI that Ilary is a human being. Whenever she turned around, her profile and especially her very bursting lips were not recognized by the computer. Having overcome this difficulty, and keeping “Ilary” frontal, in the end everything went pretty well. But, in my opinion, she is not Striscia’s masterpiece “.
Is there an even more extraordinary deepfake than the one the showgirl replies, then?
“Yes, the real masterpiece for me from a technical point of view was Jovanotti. His face has an incredible complexity both in terms of expressions and in terms of details: being a mature man Jovanotti has many wrinkles. He also has facial expressions. very rich facial and in addition he has a beard that has an infinite complexity for those who portray it. The artificial intelligence read Jovanotti’s beard as a big and long chin. We worked hard on it, it took an exaggerated amount of calculations, a week integer of calculations on multiple machines “.
Can you explain, more concretely, what kind of technology you use and how the Striscia deepfakes are born in general?
“A good deepfake requires more professionalism in different fields: from the author who decides the content to the scrupulous study of the character done by the imitator. The imitator must not only be able to imitate the voice, but must also have physical characteristics compatible with the character he We want to portray. Strengthened by this human material, we know that we have several tools on our side: clearly one is audio editing to help the voice but then, above all, the video post-production intervenes that allows us to adapt the body, the physique, the head of the person to that of the character we want to imitate “.
In short, Striscia’s deepfakes are not pure technology. There is an equally fundamental human side that we don’t see.
“Yes, absolutely. There is an unprecedented teamwork studied in detail. For example if the author exaggerates in loading the character, or if the imitator renounces micro gestures, working only on the voice, the result will be mediocre. and not truthful. Even in my work, which is to monitor the computational component, that is, in following the software in its evolution, human intervention is fundamental. If we leave everything to the machine, we will have what the computer considers the best portrait possible. But often the one who brings out the machine is an alien, because there are some characteristics that the software tends to consider salient and that instead do not return the true peculiarities of the face we want to portray. only, the computer can make mistakes. “
How does the algorithm work?
“Once we have produced the video in which we want to replace the face of the imitator with that of the celebrity, we have to prepare the model based on artificial intelligence by offering the machine some examples that we consider valid to learn how to portray that face. The advantage is in our experience. When this mass of data has been selected, the computer begins to work seriously. This type of operation requires exaggerated computational capacity and a lot of time. Each attempt that is made is called an iteration and the iterations must be hundreds of thousands to bring us to a reasonable result. We are talking about days of calculation in which the model learns, splitting itself: one part begins to portray a face and a second part of the model judges this result based on the data set that we have given it as a reference. After that, the model begins to build the face and, comparing itself almost infinitely, improves “.
After Ilary Blasi you also created Francesco Totti’s deepfake.
In the collective imagination Totti and Ilary are inseparable and it is right that they are also united in the deepfake.
When did you meet Antonio Ricci and how did your collaboration with Striscia begin?
“In 2018-2019 I was studying 3D avatars that could be used on TV. I proposed these digital avatars to Striscia, but they had another, much more powerful idea in mind: to go beyond real imitators. commissioned the first deepfake, I acted as a guinea pig by creating an imaginary “Salvini’s twin” that had my body and hair as long as me. And the result, as bad as it was, told quite well what the potential of this technique was . Then we made a second attempt with a real imitator (Claudio Lauretta): he was the first to make Salvini’s real deepfake. But the first deepfake to go on air was that of Matteo Renzi (also ‘played’ by Lauretta). deepfake caused a sensation and international newspapers talked about it: Striscia was the first television broadcast in the world to produce content using this technique, the others used ready-made content on which an they simply replaced the original face. Striscia, on the other hand, decided to use deepfake as a real production tool with lots of actors, authors and direction “.
Don’t you ever think that these deepfake videos are “too big an illusion”?
We all know the risks and have debated them for years. But the worst risk was that they were used in a less clear and declared way than as done by Striscia. Antonio Ricci explained at a press conference what deepfake was and its provocative use. The path from the technological point of view had already been traced, the question was to decide how to follow it.
What do you answer to those who pose ethical problems? Who do you think deepfakes are the result of personal data being manipulated to produce something else?
Ethical problems existed previously, digital data have already stolen many of them. Perhaps in this case we are more sensitive, because it is a type of data that involves other senses. And maybe finally we will also convince ourselves to worry and ask ourselves how to protect our data and our person.