The neural network has been taught to identify fakes on video, its accuracy being over 90%
The service can be used by media companies, news and PR agencies which seek to verify information
Russian developers who participate in the educational intensive training Archipelago 20.35 for AI specialists have created a service which recognizes fake faces technologies (deepfake) on video using neural networks.
Deepfake technologies (derived from “deep learning” and “fake”) used, for example, in mobile applications, allow for video editing by changing the appearance of people in the frame. It can include replacing the user’s face on a video with the face of a celebrity or the other way round, adding the user’s face to a scene from a famous movie. Such technologies may have generated “fake” materials, so scientists are currently looking for effective methods to identify “fake” faces. The authors of the research were able to create a technology to analyze such videos using a neural network.
“At the intensive training Archipelago 20.35 they presented a development based on a neural network, which allows to detect signs of deepfake technologies with over 90% accuracy. The project is run by the Russian company VIEN. In the future, the service will be available to media companies, news and PR agencies that need verification (fact-checking) of information, as well as ordinary users who doubt the authenticity of videos”, — told the NTI Platform.
The Fake Video Detection Service (FVDS) technology is based on computer vision models for frame-by-frame assessment and classification of objects. Using a neural network, the system evaluates the features of faces in the video and detects signs of synthetic image generation. If there are such signs, the face on the screen will be highlighted with a red frame as “fake”.
“Technologically, our solution is a customized process of trickle feeding of video content, splitting into frames, a block for applying models for Fake Video Detection and subsequent assembly into the final video,” — the press service of the NTI Platform quotes the words of VIEN Technical Director Maxim Popilin.
The authors of the technology have created and tested a prototype of a video analysis system. If they find an investor and technological partners, the developers are planning to market the service within the next two years.
About the intensive training
The educational intensive “Archipelago 20.35” was held online from 7 to 21 November. The goal of the event was to form and accelerate teams capable of creating and commercializing projects in the field of artificial intelligence and data analysis.
According to the organizers, almost 4 thousand people have taken part in the accelerator. The organizers have prepared educational programs in 26 areas of development of AI technologies for them. The top five cities in terms of the largest number of representatives at the intensive course are Moscow, Tomsk and St. Petersburg, as well as Novosibirsk and Rostov-on-Don.
The organizers of the event include the NTI Platform ANO, the Agency for Strategic Initiatives, the Innovation Promotion Fund, the Skolkovo Foundation and the NTI University “20.35”. The partners of the intensive are the Ministry of Education and Science and the Ministry of Economic Development, the Moscow Government, the Fund for Infrastructure and Educational Programs (FIOP) of Rusnano group, the Russian Quantum Center, as well as a number of universities and companies. At the end of the program 100 teams can promptly receive financial support from the Innovation Promotion Fund under the Start-CT program — up to 3 million rubles for R&D, and apply for the Skolkovo Foundation residency through a simplified procedure.
