With the rise of AI tools that take photo-editing to scary new heights, it’s getting harder and harder to determine the authenticity of digital visual data. Deepfakes challenge our eyes’ ability to distinguish between real and fake photos and videos.
But before the advent of deepfakes, and before the digital boom itself, Adobe‘s Photoshop has already revolutionized the digital photography world since its debut back in 1990.
Photoshop was and still the go-to photo-editing tool for professional and amateur creators alike. Photographers, graphic designers, advertisers, web developers, and many in other fields all use Photoshop’s photo retouching magic.
Just like it has democratized photo editing with its Photoshop software, Adobe says it wants to democratize image forensics using AI.
AI Tool Against AI Face Manipulations
Whenever a photo is altered, there would always be clues to give tampering away. While they might be invisible to the human eyes, a trained deep learning system definitely can detect them.
Aware of the implications of photo-editing technologies like its own, Adobe is focusing part of its research on developing AI solutions to “increase trust and authority in digital media.”
Last year, in a DARPA-funded research, Adobe scientists created an AI tool that spots images that were tampered by three common Photoshop techniques: splicing (combining parts from different photos into one), copy-move (cloning objects), and removing objects.
This time, a team from Adobe and UC Berkeley were sponsored by the DARPA MediFor program to design an AI to spot face manipulations. They developed an AI tool, a Convolutional Neural Network (CNN), for detecting images that were edited using Photoshop’s Face Aware Liquify feature, which allows the application of warping to human faces.
To train their AI model, the team used thousands of fake images that were automatically created by scripting Photoshop itself to apply to use Face Aware Liquify on random photos on the internet.
The team structured their research around these three questions:
- Can you create a tool that can identify manipulated faces more reliably than humans?
- Can that tool decode the specific changes made to the image?
- Can you then undo those changes to see the original?
Results show that they can answer by “yes” each of the above questions.
First, this AI tool outperforms humans at recognizing image manipulations by a significant margin. While humans were able to detect tampered faces in 53% of the cases, which is a little over the probability of chance, Adobe and UC Berkeley’s AI tool was accurate 99% of the time.
What’s more, this neural network tool can also “undo” image manipulations to reconstruct the original faces, and they successfully tested the system on real image manipulations created by a hired artist.
“It might sound impossible because there are so many variations of facial geometry possible,” says Alexei A. Efros from UC Berkeley. “But, in this case, because deep learning can look at a combination of low-level image data, such as warping artifacts, as well as higher level cues such as layout, it seems to work.”
Researchers don’t exclude the idea of a “magic universal ‘undo’ button” to revert images to their unedited version, though we’re still far from that. This is all experimental work as Adobe didn’t say there’s any commercial product coming out of this AI tool research project. But the company cited that “the journey of democratizing image forensics is just beginning.”
Comments (0)
Most Recent