So far this is a research project that they don’t want to publish in public.
Adobe, in conjunction with the University of Berkeley, has developed an algorithm that can determine if a person has been edited in Photoshop. The neural network can even suggest how to cancel the effects. This is reported on the company’s website.
Adobe is not planning to make a commercial product from the algorithm yet. However, a company spokesman told The Verge that the current experiment is just one of many attempts by Adobe to recognize manipulations with images, video, audio, and documents.
While we pride ourselves on the world’s influence of Photoshop and other Adobe products, we are also aware of the ethical impact of our technology. Fake content is a serious and increasingly urgent problem.
The study was aimed at finding the application tool Liquify. Usually it is used in Photoshop to edit the shape of the face and change emotions.
To create the system, engineers trained the neural network on a database of paired entities. One of them remained unchanged, and the second was edited using Liquify.
As a result, the algorithm learned to recognize the use of Photoshop with high accuracy. If people could recognize a fake only in 53% of cases, the neural network was 99% accurate all the time. The algorithm can also suggest how to restore the photo to its original state, but in this case, the results are mixed.
According to one of the researchers at Adobe, the idea of the Undo button, which would turn off all effects, is far from reality. He noted that in today’s world it becomes more difficult to trust digital information, therefore the tools to verify it are necessary.
As the researchers explained, the algorithm was the first of its kind and allowed to make an important step on the path to other tools. Now the company is thinking of creating neural networks that recognize changes in skin tone.