An Adobe team and UC Berkeley researchers have successfully trained an artificial intelligence system capable of tracking image manipulation through Adobe Photoshop.
Researchers hope the tool will help restore confidence in digital media at a time when deepfakes and fake faces are becoming more common and difficult to distinguish.
It would also make it possible for anyone to discover the manipulation of images. The team trained a neural network to discern images manipulated with Face-Away Liquify’s Photoshop function, which was designed to change the eyes, mouth, and other facial expressions.
When tested, the neural network successfully identified 99% of manipulated images. Compared with humans, the latter discovered manipulation only in 53% of cases. The tool was also capable of restoring images to the original format. This is not the first time Adobe has trained artificial intelligence to distinguish the modified Photoshop images, but this technique is specifically for identifying facial manipulation.
The company says we are living in a world where trust in digital information that consumes has fallen. Adobe says this is just the beginning of many efforts to come up with identifying manipulation of images.