Russian scientists have created a tool for detecting deepfake videos

[Carding 4 Carders]

Don State Technical University (DSTU, Rostov-on-Don) has created a program for recognizing video content generated using AI. The 100 KB software runs on Windows 7 and higher and detects deepfakes with acceptable accuracy.

The program is written in Python 3.11 in the Microsoft Visual Studio IDE on the Inception architecture. For face recognition, the BlazeFace model of ResNeXt and XceptionsNet neural networks was used, which were trained on the Google Cloud Platform.

The tool is easy to use: just download a suspicious video and run a check via the command line. The program will find all the frames with faces and analyze each one for signs of forgery. So, a deepfake can be given out by the tension of the lips when talking, the discrepancy between speech and facial expressions, as well as various technical nuances, up to unnaturally located pixels.

The characteristics extracted from the original are processed by a Generative Adversarial Network (GAN). There are three options for output:

• superimposed on the video verdict Fake/No Fake;
• text message indicating the probability of forgery (in percent);
• time-lapse estimates displayed on the command line indicating the degree of fake probability.

"One hundred percent result is not guaranteed: any program has some inaccuracies, - the developers emphasize. — This is due to the peculiarities of the person's facial expressions on the video, and to the characteristics of the video itself: sometimes such videos are specially made with poor quality, so that it is more difficult to identify a fake."

Currently, the creators of the anti-fake program are completing the registration of a certificate of state registration of the product, the copyright holder of which is DSTU. According to the developers, their tool will be in demand in the field of digital security, including the creators of contactless access control systems and unlocking gadgets.

"The program for detecting fake video content will become one of the modules of the future software package for countering destructive information," says the project's scientific director, Professor Larisa Cherkesova. — This system will cover all types of multimedia content on the Internet: texts-both printed and handwritten, graphic images, including photo and video files, as well as audio files."

Of the foreign developments of this plan, the most interesting is probably FakeCatcher developed by Intel — a server detector capable of recognizing deepfake video in real time with an accuracy of up to 96%. It differs from most analogs in its approach: it looks for signs that are characteristic of a person, and not for differences that give out forgery.