The photo editing capabilities offered by Photoshop are so good now, it's almost impossible to tell if an image has been altered, or "Photoshopped." So Adobe decided to remove the guesswork and developed an AI that can spot Photoshopped faces automatically.
Adobe is attempting to deal with the growing problem of fake content, which includes the use of image editing to present content that's not real. To that end, Adobe researchers Richard Zhang and Oliver Wang worked with a UC Berkeley team consisting of Sheng-Yu Wang, Dr. Andrew Owens, and Professor Alexei A. Efros to develop an AI for spotting face manipulation.
The research was sponsored by the DARPA MediFor program and used Adobe's own Photoshop Face Aware Liquify feature, which can be used to adjust and exaggerate facial features, to produce fake faces for testing. A neural network was then trained to recognize when a face had been altered in a photo using thousands of images. The images were sourced from the internet and then either passed through Liquify for alteration or given to an artist for manual enhancement.
When Adobe showed humans an original and altered version of a photo they guessed correctly which had been altered on 53 percent of the time. Once the neural network had been trained, it achieved 99 percent detection. So good is the AI that it can reverse the detected manipulations and create a very good representation of what the original face looked like.
Adobe now hopes to use the AI as part of a larger set of tools to help verify digital autheticity of images. It would also be fantastic if they could offer what the researchers term a "magic universal 'undo' button" for fake media so as to reveal the true image with a click.
