If there is absolute truth on the World Wide Web, it has to do with the frequent difficulty in distinguishing what is true from what is false. And deepfakes add to that phenomenon made up of manipulated images, fake news and invented information that often takes place on the Internet.
Derived from the merger of the expressions ” deep learning ” (research field of artificial intelligence and, more precisely, machine learning ) and ” fake ” (false), the term deepfake refers to videos and audios created by means of intelligence software artificial starting from real people’s voices, faces and bodies .
It is not just a question of simple digital fakes, nor of a game, but – when there is no consent from those involved and when the video is not expressly marked as a deepfake – a particularly serious form of identity theft, with all the legal implications that derive from it.
What is a deepfake
Creating a fake video using deep learning technology , a branch of machine learning – or machine learning – starting from real modified content: this is, in essence, the deepfake.
How does it work? This is not a simple face swap: the artificial intelligence “learns” the appearance of a real face from different angles, and then transposes it onto a target, as if it were a mask, obtaining a body with a face from another person and making it move independently, creating new facial expressions in real time and in sync with other audio.
This technique was born as an aid to cinematic special effects . We recall that Hollywood has made a real trend of transposing real and imaginary faces on other actors, managing to bring back to life those who are no longer there, as happened with the 2016 Star Wars Story film .
Initially particularly expensive and, therefore, not widespread, these techniques have begun to spread in recent years thanks to the development of apps and software that make it possible to create deepfakes using a simple smartphone. FakeApp , for example, allows anyone to try their hand at it. And so the phenomenon took hold.
And, in all of this, a decisive role was played by Reddit, a site within which, in a category created by one of its users and expressly called “deepfakes”, artfully constructed pornographic videos were published in 2017, featuring celebrities, whose facial images had been stolen.
While subsequently interrupting that user’s activity and forbidding – like other sites – the diffusion of fake porn , the spread of deepfakes has been unstoppable since then, as have the risks that derive from it when there is no explicit consent from the people whose face is featured in the faked video and when it is not clearly marked as “deepfake”.
In fact, anyone who appears in a deepfake without their knowledge suffers a loss of control over their image, as well as a deprivation of control over their ideas and positions, which can be misrepresented on the basis of false speeches and behaviors expressed in the video.
Furthermore, there is a risk that the victim will be portrayed in compromising places, contexts, situations or with people . Which – when there is no agreement with the person concerned, we emphasize – represents a threat to your privacy and dignity.
Another malicious use of deepfakes is to carry out real acts of cyberbullying against those involved, aimed at denigrating them or, in some cases, blackmailing them, asking for money or other things in exchange for not distributing the video.
In the crosshairs of those who create deepfakes there is also politics, with the diffusion of false video contents, in which not only the image of the subject in question is altered, but also his voice. The objectives, in this case, range from piloting public opinion to confusing it, up to increasing distrust in institutions and sources of information.
The deepfake case involving the former President of the United States of America Donald Trump is now well known .
And, again, the deepfake can also be used in the context of illicit telematic activities – which make use of cyber attacks such as spoofing , phishing and ransomware – to deceive people or devices and obtain data transmission from them.
Fake faces and voices also represent tools to deceive security systems based on vocal and facial biometric data, by sending – for example – deepfake video and audio messages inviting you to open messages or click on links which, according to times, they expose PCs, smartphones or other devices to illicit intrusions and the theft of sensitive information.
How to create a deepfake
We recall that an artificial neural network is a computational model composed of artificial neurons, inspired by the biological neural network. Well, each deepfake has a code, within which there is an autoencoder, which is a particular type of artificial neural network used to learn encodings of unlabelled data.
This network is trained to compress data that will later be decompressed. During the compression phase, the autoencoder tries to obtain a result as close as possible to the original image, learning to distinguish between important and secondary data during the compression phase.
What does it actually mean? That if the deep learning algorithm is fed images of cats, the neural network learns to focus only on the cat, ignoring the background . And, starting from this data, the autoencoder is able to create a cat.
The exchange of human faces works in the same way: the neural network assimilates the appearance of a person, which it is then able to regenerate autonomously.
But to exchange human faces correctly, it is necessary to be able to recognize two of them: the one that emerges from the original material and the one that is intended to be used for the exchange.
For this reason, one input (encoder) and two outputs (decoders) are set , where the encoder analyzes the input material and the two decoders each generate a different output: face A and face B.
The final effect is given by not inserting face A in the video, but face B, which does not belong to him. And here lies the difference compared to normal fakes, in which, in reality, the original face is limited to being replaced with one cut out from another image. While, in the case of deepfakes, new material is created, not just copying an image. And this to ensure that even the mimicry of the subject can trace the original face.
When we talk about the manipulation of images that portray human faces, we distinguish, in particular, four categories, namely the synthesis of the entire face, the change of identity, the manipulation of an attribute and, finally, the change of expression.
In the first case, the manipulation gives rise to human faces perfectly identical to those of real people, but completely created by the computer by means of a generative adversarial network – in English Generative Adversarial Network (GAN) – a set of methods in which two networks neural networks are trained competitively.
The change of identity, on the other hand, occurs when, in a video, the face of a person is replaced with that of another , making it impossible to recognize that it is, in fact, a fake , as happened with non-real pornographic videos featuring some celebrities, as mentioned above.
Manipulating an attribute allows you to change some aspects and characteristics of the human face (skin color, hair color, gender and age), as happens when using FaceApp . While, the change of expression allows, for example, to add a smile to the face of someone who appears serious or sad in a photograph. From this, it is possible to understand the danger deriving from this type of manipulation, through which it becomes possible to create from scratch the lip movements related to a speech and transpose them on the face of another person, modifying their voice.
What are they used for
We have already mentioned the risks related to deepfakes when a person – without their knowledge or consent – becomes the protagonist of a video.
Now, however, we focus on the ethical uses of this practice. Starting with marketing, where, in particular, in the case of advertising and communication campaigns aimed at more foreign markets, the language barrier can be a problematic aspect in the creation of videos.
Here then is that, by intervening – manipulating them – on the expressions of the faces of the protagonists, the movements of the lips can be adapted to other languages, thus obtaining a good result, valid and effective in penetrating new international markets.
Another example, always referring to the marketing sector, concerns the adaptation of advertisements and online advertising campaigns according to the different reference targets. The emblematic case is that of the promotional video which has a Western man as the protagonist . Video that – rethought for the Asian market – deepfake techniques are able to transform, attributing oriental facial features to the subject .
But the deepfake can also prove to be particularly useful in another area, crossing that same privacy which, when there is no complicity of the interested parties, is harmed. We are referring to those videos in which, thanks to deepfake techniques, the identities of characters are masked – for example activists – whose choices, positions and statements are not allowed within certain socio-political scenarios.
This is the case of the documentary Welcome to Chechnya on the persecution of LGBTQ individuals ( Italian acronym of: Lesbica, Gay, Bisexual and Transgender) in the Russian republic , the first film where the deepfake is used to protect the identities of activists fighting persecution.
Finally, another example of “good and useful” use of the deepfake is linked to didactics and games with educational purposes, reviving – within videos that portray them today – historical figures, celebrities from the history of art , literature or music.
Some examples of famous deepfakes
As mentioned, deepfakes were born on Reddit by a user who, in December 2017, posted porn movies in which the female protagonists have the faces of some Hollywood stars (unaware of this), including the actress Gal Gadot – made famous thanks to Wonder Woman – Jessica Alba, Taylor Swift, Daisy Ridley and many others.
With the story on Reddit closed, it was Barack Obama, in 2018, who became the protagonist of what is considered, to date, the most famous video dedicated to deepfakes, in which the former president of the United States, indistinguishable from the real one, with his voice and perfectly synchronized lip, he makes serious and not without foul language statements, going so far as to call Donald Trump a “complete asshole”.
With nearly 7 million views, the video in question was created for educational purposes, to understand the extent of deepfakes and warn about its risks.
Some time later, the first real victim – as unaware of what happened – is Nancy Pelosi, speaker of the United States House of Representatives, who, in a manipulated video from May 2019, appears as if she were speaking at a completely drunk conference . Video so perfect and credible as to convince many, including the former mayor of New York Rudy Giuliani.
In response to Facebook’s refusal to remove Pelosi’s video, it’s the turn of Mark Zuckerberg, the victim of a video in which he boasts of “having control over our lives”.
From deepfakes made in the USA to our own, with a famous counterfeit video in which Matteo Renzi claims that “Conte has the face of an idiot and Zingaretti has the charisma of Bombolo” and raspberries at President Mattarella.
Today, the phenomenon is so widespread and known to the general public that on TikTok there is even an account entirely dedicated to Tom Cruise deepfakes, with a whole series of video contents that show the actor playing golf, doing a trick of magic or in mundane situations, such as washing oneself. But – in this case – the account description clearly warns that it is a fake and, more precisely, a “parody”.
Another famous example of a deepfake featuring, this time, a famous artist of the past, is the one by an advertising agency that has resurrected Salvador Dalì in the role of unpublished guest of the Dalí Museum in Florida. The fake video in honor of the Catalan master was created by extracting more than six thousand frames from his old interviews, then processed using sophisticated machine learning techniques.
How to recognize a deepfake
There are some elements on which to focus carefully in order to identify deepfake manipulations. First of all, the fact that the image can appear a bit grainy or blurry (“pixelated” is the technical term), with the illumination of the face changing from that of the surrounding environment.
Another detail is given by the eyes of the protagonists of the videos – which could move in an unnatural way – and by the mouth, which could appear deformed or too big while the subject speaks.
With regard, specifically, to the movement of the eyes, it is the iris and the movements of the pupils that must be observed carefully. In many deepfakes, in fact, the movement of the muscles around the eyelids does not correspond to that of the eyeball.
For this reason, experts urge you to watch videos that appear to be counterfeit on a large screen, and not using a simple smartphone.
It should be added that large digital companies have for some time been studying methodologies aimed at recognizing deepfakes, including ad hoc artificial intelligence algorithms and punctual systems for reporting by users.
In any case, it is essential to remember that, if there is any doubt that a video or audio is deepfake made without the knowledge of the interested party, it is necessary to avoid sharing them and decide to report them as “false” to the platform that hosts them.
If, then, it is believed that the deepfake in question commits a crime or a violation of privacy, one can directly contact the Police Authorities (for example, the Postal Police) or the Guarantor for the protection of personal data.
What to do to defend yourself
The main tool of defense against deepfakes is represented by everyone’s attention and responsibility, as underlined by the Privacy Guarantor , who recommends, first of all, to avoid spreading personal images or images of loved ones on the net in an uncontrolled way.
And, with regard to deepfakes against companies and organizations, the Guarantor underlines the importance – in addition to awareness and training – of concrete organizational and procedural measures that guide and help personnel not to be deceived .
In this sense, the episode which, in March 2019, saw the CEO of an English company send a large sum of money to a Hungarian bank account following a phone call “apparently” from the CEO of the parent company, whose voice had been perfectly cloned.
This is an example of how, to the awareness of personnel at all levels, it is now necessary to add a precise security plan which identifies, in the company, those data and operations at risk and which foresees – at all company levels, including highest – a double authorization or unambiguous and predetermined procedures for imparting the different provisions.
We recall that the GDPR – General Data Protection Regulation – which became operational on 25 May 2018 – made it necessary to extend this practice to all personal data processed, thus including a large part of the company’s business. And, given the speed with which they are evolving and spreading, the risks associated with the existence of technologies based on artificial intelligence can no longer be ignored or overlooked .
Those organizations that have long established compliance with the legislation on the processing of personal data will only have to expand the register of processing – including non-personal data – identify the phases in which an attack based on deepfake technology could be launched and put in place control in line with the risk of a possible violation.
