Changing faces virtually is the order of the day, and the infinity of news about deepfakes that come to us day after day attest. As a proposal to try to solve this problem, Adobe has worked together with the UC (University of California) Berkeley on the detection of manipulations on faces through Photoshop.
This proposal, which, for the time being, does not fit into their business plans, is according to them one of the company’s moral commitments, which claims to be working on ways to detect when an image, video, audio or document has been manipulated. According to Adobe, fake content is a problem, and artificial intelligence can help detect it.
A powerful tool that not only detects changes, but also reverses them
From the blog of Adobe tell that his own team, together with collaborators at UC Berkeley, has managed to develop a method to detect changes in photographs. The main objective is to create a solution capable of identifying faces modified with greater precision than a human being, of decoding these same changes and of reversing them by itself, without editing by the user of Photoshop.
For this, they have focused on the Photoshop “Liquify” tool, which is frequently used in the world of photography and modeling to make deep modifications in faces, since with this tool we can change the appearance of a face almost to the full.
“We started by showing pairs of images (an original and an alteration) to people who knew that one of the faces was altered.” For this approach to be useful, it should be able to perform significantly better than the human eye in identifying edited faces. “
The process? Train an artificial intelligence that compared thousands of images of the internet without modifying images synthetically modified with this tool. In the same way, an artist intervened in the modification of images, to add the human component to said modifications. The resulting algorithm was able to recognize the modified images 99% of the time, while the human test subjects barely exceeded 50% accuracy rate. Going further, this AI is even capable of reversing the images to their original state, after analyzing the aspects that have been modified.
From Adobe they say that we are still far from a magic button that reverses all the changes in a modified image, but that its tool is a good point to mark this path, and raise awareness about how easy it is to manipulate an image and that the human being see no account of such manipulation.
Without going any further, recently a controversy arose with a video of Nancy Pelosi, president of the House of Representatives in the United States, in which apparently appeared drunk, and that turned out to be false. Similarly, a few days ago how a video with a deepfake Mark Zuckerberg began to spread through the network was not removed from social networks, since Facebook does not have a policy stipulating that the information published in the network must be true.
These cases put on the table the need for precise and fast tools that detect the false content, one of the main purposes that Adobe seems to transmit, and on which they will base future tools.