Deepfakes: how and for what purpose are they created and how to detect them?

Home /

Blog

16.11.2023

Internet

15509

You've probably come across deepfakes on the Internet — photos and videos that look very similar to real ones, but show people saying or doing things they didn't actually do. The term "deepfake" is a combination of two concepts: "deep learning" and "fake". That is, they are photos or videos artificially generated using the knowledge gained in the process of neural network training. What tools are used to create them, why and how to distinguish fake images from real ones? Let's understand together.

Special applications for creating deepfakes

Today, anyone who has a smartphone can download one or another app designed for creating deepfakes. Here are a few of them:

Reface

This is the most popular App Store application, moreover, created by Ukrainian developers. Using the online service, you can replace one face with another in videos or GIF images. All you need to do is take selfies and select the target video, and the service will create an animated picture. But the accuracy of the final result depends on the symmetry of the face in the photo and the quality of the incoming data.

Avatarify

The Deepfake app, available on iOS and Android, allows you to upload any photo from your smartphone's gallery. By transferring emotions to the face, a neural network brings it to life and saves the result as a short video.

Wombo

The service gives static faces of people in the photo a singing facial expression. To do this, the user only needs to take a selfie and choose a music track from the list of suggested tracks, and then the system animates the face and synchronises its lips with the selected song.

MyHeritage

This free app has a "deep nostalgia" feature that animates old photos, giving them new life. All you have to do is upload an image and tap the animation button.

DeepFakes Web

Need to replace one person's face in a video with another? DeepFakes Web can help you with that. You only need to upload the source and target videos, and then the video will be processed in the cloud within 30 minutes. The app is available for download on both iOS and Android.

Harms and Benefits of Deepfakes

The availability of deepfake technology facilitates the spread of multiple manipulations in society. By passing off lies as truth, it is easy to persuade people to believe in something that doesn’t really exist. And this is very dangerous and can have significant consequences: destruction of reputation, destruction of marriage, disruption of elections, etc.

Deepfakes became most popular with the advent of the FaceApp app in 2017. Today, the created deepfakes are being posted en masse on the Internet network. In particular, deepfake videos can be found on YouTube. For example, like this popular video in which Barack Obama seems to be talking offensively about Donald Trump:

The use of the results of deepfake technology can harm the reputation of a particular person. Thus, one of the first and most famous deepfakes in history was a video with actress Gal Gadot: in the video, the face of a film star was "glued" to the body of a porn actress. And later other actresses (Natalie Portman, Emma Watson, Natalie Dormer, etc.) became victims of similar manipulations.

Falsification of data and their further presentation as evidence in court is another unfair way of using photo and video deepfakes. As well as deception of facial or voice recognition systems by imitating biometric data in order to violate a person's personal security.

But at the same time, the technology is also being used to its advantage when performing work tasks. For example, in the film industry, when you need to replace an actor, correct mistakes in filming, synchronise the movement of actors' lips and audio tracks when dubbing films in different languages, etc. Or in marketing, as the Lays brand did in its advertising campaign, in which it used the deepfake of star footballer Lionel Messi:

Another example of eco-friendly use of deepfakes: a documentary film about American chef Anthony Bourdain used a spoofed celebrity voice to read a desperate personal email.

Also, the results of neural networks' "creativity" are becoming a popular way of entertainment. One of them is the publication by users in social networks Instagram and TikTok of excerpts of popular films with replaced faces of characters created through special applications. Here we can also recall the videos posted on the @deeptomcruise account in TikTok as if with Tom Cruise, which gained millions of views in three days. From this we can conclude that the account of every public person should be verified, so that it would make it easier to identify fake pages.

Learning to recognise deepfakes on the web

While it was easier to recognise created deepfakes at the stage of technology's emergence, as it develops, it becomes more and more difficult to do so. Special software can help to identify fake photos and videos with a higher probability. But practice confirms that the best results can be obtained on the examples of celebrities, because in free access there are many materials on which artificial intelligence has the opportunity to train. It is more difficult for machine algorithms to cope with the exposure of deepfakes with ordinary people.

However, it is still possible to detect a fake, paying attention to such discrepancies:

Visual. Remember how the first AI-generated images of humans had strange palms with six fingers on them? Of course, modern deepfakes aren't so goofy anymore. Now you should be alerted to unusual body movements or facial expressions on a person's face (e.g., unblinking eyes, inconsistent shadows or lighting, etc.).

Audible. If you notice that the character's pronunciation on the video is not synchronised with the movement of his lips, this also clearly indicates a fake.

Contextual. Don't forget to engage your critical thinking, which will help you answer the question, "Is the character in the video even able to say and do what I see?".

Summary

The more advanced machine learning algorithms become, the better their performance becomes, and thus the more realistic the deepfakes become. But it is still possible to distinguish neural network-generated photos and videos from human-generated ones by certain attributes. And it's a good thing when it prevents synthesised human-generated images and videos from being used as a weapon to spread misinformation that could seriously harm others.