Adobe is working with UC Berkeley scientists on a method to identify manipulations in portraits. This includes photo edits made with the Face Aware Liquify tool in Photoshop. Such manipulation could not only be discovered by the AI, but could also be undone with it.
The researchers have published a study describing a method that detects manipulations applied to images of people’s faces. The technology consists of a neural network that has been trained with manipulated images that have been created using scripts with Photoshop and the Face Aware Liquify tool.
To test the assessment skills of the AI, the researchers used pairs of images with an original, unprocessed image and a modified photo. The scientists filled the system with thousands of photos from the internet, changing various facial features. The neural network managed to pick out manipulated photos with an accuracy of no less than 99 percent. If people left the assessment to people, they appeared to choose the correct picture in only 53%.
Moreover, the technology proved to be able to determine the specific areas of a distorted face and which methods were used for this. It could also calculate the best way to reverse the edits to restore the photo to its original state.
According to Adobe, trust in a world where image editing is omnipresent is becoming increasingly important: “Fake content is a serious and increasingly pressing problem. Adobe is committed to finding the most useful and responsible ways to bring new technologies to life – and constantly looking for new technologies, such as artificial intelligence (AI), to increase trust and authority in digital media. ”