Facebook and Microsoft summed up the results of the contest against dipakai

Companies Facebook and Microsoft summed up Deepfake Detection Challenge — a competition for developers, which aims to create solutions to combat the technology of substitution of the individual in the video. The developers managed to achieve a recognition accuracy of over 82 percent on a standard test dataset and slightly more than 65 per cent datasite complicated: use distracting components (e.g., filters, or labels). Read more about the results of the project described in the blog on the website Facebook.

Technology of creation of diphenol in recent years develop very quickly, and after that there is a need for solutions that can effectively help to recognize them. To help in this decided Facebook and Microsoft in September last year, they announced a contest for developers Deepfake Detection Challenge, on the idea that any third-party developers will be able to propose and create a model that can determine the substitution of the persons in the video. For developers, the organizers promised to create dataset for which hired professional actors: so it will not be used for user data.

The contest was attended by 2114 developers who have created more than 35 thousand models. Evaluation of the effectiveness of the algorithms was carried out in two ways: the first used the developers test dataset, and the second — closed and complicated (it was used live with running lines, filters, and actors who partially covered his face). According to the tournament table on Kaggle, the assessment of work using a standard test dataset won user under the name Good At Curve Fitting: the accuracy in determining diphenol using its algorithm made 82,56 percent.

When used to check closed dataset defeated Belarusian developer Selim Seferbekov of the company Mapbox: his algorithm was able to identify deepface with precision 65,18 percent (standard inspection he also took the fourth place). All developers who took the prizes used in the algorithms pridorozhnoy a kind of convolutional neural networks EfficientNet that in a recent Preprint described by the developers of Google Research.

Leave a Reply

Your email address will not be published.