Facebook has developed an AI algorithm that prevents AI from recognizing faces in videos

Facebook AI Research claims to have created a machine learning system to avoid identifying people in videos. Startups like ID and a number of previous ones have already created similar technologies for photographs, but for the first time the technology allows you to work with video. In the first tests, the method was able to disrupt the work of modern face recognition systems based on the same machine learning.

AI for automatic video modification does not require additional training for a specific video. The algorithm replaces a person's face with a slightly distorted version to make it harder to identify using facial recognition technologies. How it works - you can see in demo video.

“Face recognition can lead to a loss of privacy, and face replacement technology can be used to create misleading videos,” the document explaining the approach says. - Recent world events related to the progress and abuse of facial recognition technology make it necessary to understand the methods that successfully deal with de-identification. Our method is so far the only one that is suitable for video, including broadcasts, and provides a quality that is far superior to the methods described in the literature.”

Facebook's approach combines an adversarial autoencoder with a neural network. As part of the training, the researchers tried to trick neural networks trained to recognize faces, Facebook AI research engineer and Tel Aviv University professor Lior Wolf told VentureBeat about this by phone.

“In this way, the autoencoder is trying to make life difficult for a neural network trained on face recognition, and this is actually a universal method that can also be used if it is necessary to develop a method for masking speech or online behavior or any other type of identifiable information that needs to be removed,” he noted.

AI uses an encoder-decoder architecture to generate distorted and undistorted images of a person's face, which can then be embedded into a video. Currently, Facebook has no plans to use this technology in any of its applications, a representative of the social network said in an interview with VentureBeat. But such methods can ensure the creation of materials that remain recognizable to humans, but not to artificial intelligence systems.

Facebook is currently facing a $35 billion lawsuit related to the social network's automatic facial recognition issue.

Source: 3dnews.ru