As the flagship editor of Photoshop, the reference for image editing, Adobe now offers to identify modified photos, and it is the artificial intelligence that looks for changes made to an image.
The artificial intelligence fundamentally changes our relationship to the image, especially with the appearance of deepfake for creating false images indistinguishable from reality. However, a very particular software had already made itself known for the alterations that it allows on the images, almost impossible to detect when they are well realized. The Photoshop software has been so successful that it has even passed into everyday language, giving rise to the verb “photoshopped”.
The Adobe publisher has decided to fight against false images and has created a team that works on the issue, with the collaboration of researchers from the University of California at Berkeley. Their research bears especially on the use of artificial intelligence techniques and deep learning to detect specific signs such artefacts left by the software. Adobe had already used artificial intelligence to detect changes made by its software, especially when images are connected, or parts cloned or replaced. This time, the study focuses on the tool “Modification of facial features”.
A neural network that identifies 99% of the images faked
Photoshop software is able to automatically detect different facial features and propose changes using a simple cursor, such as the shape of the face, nose, the spacing or the size of the eyes, etc. This tool is therefore within the reach of all users. The researchers relied on a convolutional neural network (CNN), driven on pairs of images, with the original and a version treated with Photoshop. The training material includes thousands of images taken on the Internet and then edited automatically with the Photoshop tool, but also images modified by an artist for more diversity.
The researchers also presented pairs of images to human volunteers, who after a training phase attempted to detect the altered images. Their success rate was only 53%, barely better than chance. The artificial intelligence created by the team was able to detect up to 99% of photoshopped images.
The IA even manages to cancel the changes
The tool does not stop when changes are detected. Thanks to the information on the tool used, and the training on the pairs of photos, with the original and the modified version, the AI is able to estimate the appearance of the starting image.
The algorithm calculates the effects of the tool and recreates the original image, with a precision that even astonished the researchers, thus creating a real “cancel” function, which can be used on any image altered with the Photoshop ‘Face Modification Tool’.
Adobe says it wants to democratize the forensic analysis of images. One of the firm’s researchers, Richard Zhang, said, ” The idea of magic and a universal cancel button to remove the changes on the images is still far from reality. But we live in a world where it becomes more and more difficult to trust the digital information we consume, and I look forward to continuing to explore this field of research. “