The press service of Sberbank reported that the company received a number of patents for technologies for recognizing deepfakes. Efficiency rates on independent test data were 98%.

In a press release, the company said that it received patents for a number of ensembles of neural network models of the EfficientNet class (patent No. 2768797) and a method for amplifying and analyzing microchanges in the colors of objects on frames using AI (patent No. 2774624). A system incorporating these technologies helps to accurately distinguish deepfake videos.

The principle of the system is to analyze faces in videos. With this method, the faces are divided into several frames and the red color of the RGB palette is extracted for each frame. After that, an array of red color changes is compiled for each frame pixel and the change in this spectrum is calculated. Allows artificial intelligence to determine whether the image was created artificially.

Sber constantly improves its cybersecurity systems and conducts scientific research in this area. Our patented inventions will effectively solve the problem of detecting information attacks and cyber attacks using deepfake technologies. Synthetic videos used to attack Sberbank and our clients will be detected at the earliest stages.

Deputy Chairman of the Board of Sberbank Stanislav Kuznetsov

It is noted that the system can be used to analyze videos with several faces in the frame, which significantly increases the efficiency of the method. The overall success rate on independent data was 98%. Representatives of the company said that the development can be used to protect against cyber attacks bypassing the Face Recognition and Liveness Detection systems, to combat fake news and protect video conferencing negotiations.

bbabo.Net