Photoshop can take a bad picture and make it look better. But it can also take a picture of you smiling and make it into a picture of your frowning. Altering images and video can of course be benign, but it can also have nefarious purposes. Adobe teamed up with researchers at Berkeley to see if they could teach a computer to detect a very specific type of photo manipulation. As it turns out, they could.
Using a Photoshop feature called face-aware liquify, slightly more than half of the people tested could tell which picture was the original and which was retouched to alter the facial expression. However, after sufficient training, Adobe’s neural network could solve the puzzle correctly 99% of the time.
It might seem odd to focus on that specific type of edit, but it is useful for making very subtle changes to a person’s face. Earlier research worked on detecting cruder manipulations.
It sounds as though the neural network could determine which of the two photos was altered. This seems like an easier problem then simply identifying a picture as altered without another photo to compare. That would be a lot more useful, but also probably a lot more difficult, as well.
We suppose neural networks detecting fake photos is no more outlandish than asking them to judge our photography. We’ve even seen them correct for depth of field.