Check, don’t trust: to distinguish a video from a deepfake

Apple has introduced a neural network called NeuMan, capable of generating deepfake videos based on existing videos. A 10-second video with your participation is enough for a new neural network to make a fake.

How to create a fake?

To date, there are many projects and applications that are freely and free of charge, with which you can create deepfakes. You don’t need to be able to program for this.

ZAO is an application for Android and iOS. The essence is to modulate the voice of celebrities and superimpose their face on the actor’s body. The application is based on the work of trained neural networks. They do not just replace one image with another, but also copy the facial expressions and movements of the hero. However, there are drawbacks – the program is tailored to China, so the processing of European-type persons has not yet been worked out.

DeepFace Lab is one of the most popular programs for creating deepfakes. The project repository has almost 8 thousand forks and 34 thousand stars on GitHub. The system is designed simultaneously for users without knowledge of deep learning frameworks and for developers.

Reface is a startup that is transitioning from a face–changing app to a social platform. The project is aimed at creating deepfakes for the fashion and advertising industries. One full-face photo is enough to see your face in popular GIFs, movie excerpts, advertising or a game. Developers simultaneously with their services create antidotes to help recognize the generated images and videos.

How to identify a fake and not become a victim of fraud?

You can try to determine the deepfake both independently and with the help of special programs.

Independent search

Deepfakes have distinctive features from the original — they may have artifacts that are perceptible to the human eye, or small deviations. They will also be warning signals that you are facing a fake.

The table below contains some characteristic features for deepfakes.


When it is not possible to find a deepfake on your own, professional expertise will help. So it happened with the alleged deepfake of the President of Gabon.

In the fall of 2018, a tense political situation was formed in Gabon due to the long absence of President Ali Bongo in the country and rumors about his death. Amid speculation, the president’s advisers promised that he would deliver his usual New Year’s address. As a result, the government published a video that raised more questions than answers from the people. Some Gabonese have come to the conclusion that the video is a fake. The chief technologist of McAfee and his cyber defense team conducted two examinations and gave an almost 92 percent probability of authenticity of the video message.

Algorithms and programs that find deepfakes

Ironically, artificial intelligence helps not only to create, but also to identify deepfakes. However, many existing detection systems work qualitatively only on photos of celebrities due to the fact that neural networks train for hours on freely available frames.

Algorithms already exist to help creators verify the authenticity of their videos. For example, a cryptographic method is used to insert hashes at certain intervals throughout the movie. The hashes will change if the video is modified.

Programs that can be used to detect deepfakes:

  • Deepware is a free, open source video scanner and Android app.
  • Image Edited – the program determines for free whether changes have been made to the snapshot. If there were, it tells you in which graphic editor they were made. The service analyzes photos at the pixel level and their colors.

The above programs work only for video and photo fakes. Audio deepfakes are more difficult to calculate, and their creation is less labor-intensive and does not require large capacities.

Deepfake technology will develop, as well as methods of its detection. To date, it will not be difficult to detect a deepfake video. Many examples of deepfakes are currently just funny parodies or experiments designed to explore the limits of deep learning technology. However, the situation is different with audio dipfects, they are already quite realistic today. It will not be difficult for experienced attackers to register a dialog script. In addition, there are programs that create audio deepfakes online.
At the level of organizations, there should be clear regulations for confirming financial transactions or other important actions. It is necessary to prescribe instructions according to which the performing personnel will receive confirmation of their actions either through another communication channel or through a second proxy.

Based on VC materials

Leave a Reply

Your email address will not be published. Required fields are marked *