Neural network learned to recognize fake images

With the development of neural networks, the performance quality of fake images has reached a whole new level. It is almost impossible to distinguish a fake from the original - at least without the help of a neural network. This is a two-tier software tool introduced by Californian scientists. He is able to determine the deepfake-image created by "colleagues in the shop," in minutes.

Scientists at the University of California have developed an AI algorithm capable of determining, with an accuracy of up to 95%, whether the original image has been processed. According to the developers, this tool will be actively used to detect fakes on the Internet. To train the network, she was given a thousand original images and their fakes. One of its components was a kind of so-called “recurrent neural network”, which breaks the image into small fragments and views them pixel-by-pixel. The other part passes the entire image through a series of coding filters that allow it to view the image at more consistent levels.

As an example, the developers gave an image of a bird that was superimposed on a photo of an empty branch of a tree. The pixel algorithm in this case marks the pixels around the bird's claws as problematic, while the encoder algorithm can identify patterns in the summary image. As long as both components of the neural network mark the same image area around the bird, the algorithm will classify the photo of the bird and the branch as a potential fake.

Nevertheless, according to the developers, even the recognition efficiency of 99% does not guarantee the "reliability" of AI when recognizing fake images - therefore, one cannot speak about victory over "dip-fakes" yet.

4pda.ru