San Francisco (AFP)
Scientists from Facebook on Wednesday presented a method that should allow, thanks to artificial intelligence (AI), to flush out the "deepfakes", these hyper realistic fake images, as well as to determine their origin.
The "deepfakes" are problematic on the internet because they can be used to manipulate internet users or to defame, by making people say or do things that they have not said or done.
These arrangements are based on artificial intelligence technologies.
"Our system will make it easier to detect deeepfakes and track related information," said Tal Hassner and Xi Yin, two social network researchers who have worked on the subject with Michigan State University, in a statement.
Their method should provide "tools to better investigate incidents of coordinated disinformation that use deepfakes," they said.
To develop their system, they used the technique known as "reverse engineering", which consists in deconstructing the manufacture of a product or, in this case, a video or a photo.
Their software identifies imperfections added to the editing, which alter the fingerprint of the images.
In photography, this fingerprint identifies the camera model used.
In computer science, "it can be used to identify the generation system that was used to produce the trick", explain the scientists.
Microsoft introduced software last year that can help spot photo or video deepfakes, one of several programs designed to combat disinformation ahead of the US presidential election.
At the end of 2019, Google had made public thousands of video deepfakes made by its teams to make them available to researchers who want to develop methods for detecting manipulated images.
© 2021 AFP