How to detect fake profile photos on the networks?

In the era of deepfake, suspicious-looking accounts with often nauseating comments abound on Twitter. Trolls who often do not hide behind the image of an avatar. Well, not really.

Dare ugly, beautiful, big bushy or big necks (everyone has their own references). So many ordinarily banal faces parade under the indifference of our thumb which caresses, or mistreats, the touchpad of our smartphone. And then a comment annoys us. A Pavlovian reflex invites us to click on the identifier of the profile of the author of the intellectual attack with 280 characters. A face like a thousand others but a dubious ultra-bright smile and an account only in its infancy…suspicious. Are we right to be suspicious? Certainly !

Pinned by Victor Baissent, specialist in tech and web issues, in a long thread posted a year agothe use of photos generated by artificial intelligence has become commonplace. The debate polluter doesn’t impersonate, a market of free faces awaits him on sites created for this purpose, like ThisPersonDoesNotExist.

The technique used is that of GAN, for Generative Adversarial Networks, it is part of Machine Learning. It is a machine which, like a studious pupil on the benches of the school, learns and which will, thanks to this learning, improve and gain in precision. In short, a maker of realistic images. But the little genius is not without flaws.

On the same subjectThierry Ardisson interviews dead celebrities: the return to grace of deepfake, at what cost?

“To save the Internet, you have to start by not suffering from it”: can we still save the web?

Imprecise artificial intelligence

The Twitter account @_The dwarf_ supplemented the statements of the expert. Spotting fakes is easier than you think.

Rather scary fails. ThisPersonDoesNotExist

Minnow eyes, uneven skin quality, unclear or mottled teeth, aura that surrounds an asymmetrical face, somewhat pixelated hair, you don’t need a lynx’s eye to distinguish these abnormalities.

The framing is not left out. A close shot, from the front, without any element that conceals the features (no hand on the cheek, scarf or mask) and a blurred or neutral background. Sometimes even a very artistic blur, even whimsical. The AI ​​becomes a monster simulator, the precision of details is its worst enemy. Don’t put earrings or glasses in his way, he won’t digest them. The most obvious clue concerns the position of the gaze: the latter will always be aligned according to a fixed and immutable grid, and therefore always positioned in the same place.

The game goes further, a lucrative trade has been established in the United States. For the modest sum of $2.99, a company will provide you with a figure of the gender, age and ethnicity of your choice. It would take a magnifying glass to find the failure. The next Tinder date may turn sour…


🚨 WAIT ! You liked this article, and you want to support us?

💡Log in for free to your NEON space. It allows you to be alerted of our next articles on the same subject (you can click on “follow this subject” at the top of this page), to subscribe to our newsletters (a weekly best-of and our sex NL) and save your articles for later. In accordance with the GDPR, we don’t sell your data, and we won’t spam you, we promise. It may be a detail for you, but for us it means a lot (and it keeps us free, by the way).


How to detect fake profile photos on the networks?