Adobe builds AI to recognize face manipulation in Photoshop

Adobe has developed a way to recognize via software whether facial features in a photo have been manipulated. This works with photos that have been edited in Photoshop, and it is also possible to revert the edits.

Adobe explains how that works in a blog post. Adobe’s research department collaborated with scientists at UC Berkeley on the tool. The research team has set up a neural network that allows photos of faces to be scanned for edits. In addition, the photos in question must have been post-processed in Photoshop with the Face-Aware Liquify tool.

The researchers collected thousands of photos from the Internet, in which a number of randomly chosen photos were post-processed by an artist, manipulating various facial features. After the training, the developed neural network was presented with two versions of a photo, an original version and a manipulated one. The manipulated version had to be selected.

It turned out that in 99 percent of the cases, Adobe’s AI managed to fish out the manipulated photo. When that same test was done with human raters, it turned out to be only 53 percent of the time; slightly higher than if a choice were made at random.

Adobe states that it developed the tool because of the increasing focus on the problems caused by manipulated photos. In the past, the software company has come up with a neural network to recognize manipulations performed with other Photoshop tools.