Adobe-AI recognizes manipulation of faces in Photoshop

Scientists from Adobe and UC Berkeley have developed an Artificial Intelligence (AI) that can detect changes to images made with Face Aware Liquify in Photoshop. The operations can also be undone with the AI.

Adobe Photoshop was launched in 1990 and has since then allowed users to customize numerous photos. However, Adobe states that this technology also has ethical implications. Photoshop also enables fake content that is indistinguishable from real content. The new AI function should make the technology more responsible.

New network

The AI was developed by Richard Zhang and Oliver Wang of Adobe and Sheng-Yu Wang, Dr. Andrew Owens and Professor Alexei A. Efros of UC Berkely. The AI specifically focuses on adjustments made with the Face Aware Liquify feature, as it is popular for adjusting facial features. It’s also about adjusting facial expressions. To recognize the modifications, a neural network was trained.

The team created a set of training images for that neural network, by modifying Photoshop’s thousands of images from the Internet with Face Aware Liquify. Then some of the pictures were randomly chosen for the training. In addition, an artist was hired to adapt images that were in the dataset. By adding human creativity, the changes and techniques used became broader, according to Adobe.

The test consisted of showing two images – one original and one modified – to people who knew that one of the images had been changed. In 53 percent of the cases, people were able to recognize the modified image. The neural network did this in 99 percent of the cases. In addition, the tool recognised specific areas and methods where the face had changed. This allowed it to restore images to their original state.

Future

What exactly will happen to the technology is not yet clear. Adobe has not said anything about the future of the AI tool. However, the company says that it continues to explore content authenticity with its customers, partners and community.