Bbabo NET

Science & Technology News

Pay attention to faces: how to recognize deepfakes

The popular rapper in the video promises to give up to 100,000 rubles to everyone who follows the link: “Three hours left. Take your 100,000 rubles as soon as possible. In fact, this video is a deepfake, in other words, a fake created using artificial intelligence and neural networks. As a rule, these videos are not of the highest quality, but even they can be misleading. When Channel4 aired a deepfake of Queen Elizabeth II dancing on a table, some viewers believed it.

Deepfakes are often used as a harmless prank, such as replacing an actor's face with their own, but it happens that criminals use porn deepfakes for blackmail and extortion. The number of deepfakes is growing rapidly. If in December 2019, Sensity AI detected about 24,000 fake videos, then by the end of 2020, there are already 85,047 fake videos.

How to recognize a fake?

Various applications are used to create deepfake videos, for example, RefaceAI (more than 100 million downloads in the Google Store), or more advanced programs with ready-made libraries, such as DeepFaceLab and SimSwap. Some of the applications require minimal technical skills and allow you to create a deepfake video based on just one photo.

“Even high-quality deepfakes used in the film industry and advertising still lack naturalness or realism,” notes Vladimir Kalugin, Operations Director of Digital Risk Protection Group-IB. - The slightest unnaturalness in the appearance and behavior of the object causes anxiety and dislike among the audience - this is the so-called "uncanny valley" effect. Therefore, when analyzing a suspicious video, first of all, you should pay attention to the main object of the overlay - the face.

A deepfake can reveal many signs:

excessive pixelation, defects, fuzzy and blurry image, duplication of elements;

unnatural facial expressions, especially when blinking, moving eyebrows and lips;

low video quality, behind which they often try to hide the incorrect operation of neural networks;

differences in physique, hairstyle and voice from the original. Attackers often select actors of similar type, and the voice can also be adjusted using special technology, for example, in the RealTimeVoiceCLoning, Resemble and Descript programs;

in the most popular solutions, only the face is replaced, so you can see the border of the face overlay, the difference in shadows, lighting, and skin tone.

Let's move on from theory to practice, let's analyze the deepfake with "Elon Musk":

More importantly, fraudulent schemes often use not deepfakes, but real fragments of archived videos with famous people. For example, many stars have participated in various advertising or marketing campaigns to promote any product. Such videos can only be edited and re-voiced by Internet scammers - and there will be no technical manipulations with the imposition of faces.

Pay attention to faces: how to recognize deepfakes