Technology can and must make the world a better place, but some of the tools we use to make positive changes can also be returned to us to promote lies. Adobe hopes to help slow the spread of false information with tools that allow detect manipulated images. The company worked with UC Berkeley scientists to develop a neural network that can detect false photos and go back to restore their original appearance.
Adobe already has some features in Photoshop to prevent users from using them in a shady way. For example, image recognition tools block the manipulation of banknotes. These files will not even be open. Of course, you can not prevent people from opening files containing faces, but the new proof-of-concept tool from Adobe and UC Berkeley could tell you if these faces have been altered.
The tool detects the use of Adobe's Aware Liquify Tool, which facilitates the subtle alteration of a person's face on an image. The detection is based on a convoluted neural network formed with the help of examples of head shots published with Photoshop as well as the same photos before modification.
Before testing the accuracy of the neural network, Adobe wanted to know if humans were good (or terrible) to detect manipulated photos. He showed the altered and unaltered headshots to human volunteers, asking them to determine which photo was wrong. Unfortunately for humanity, they were able to choose the fictional photo only 53% of the time, the two images in front of them. The artificial intelligence formed at manipulated photos was able to spot the fake with an accuracy of 99%.
The detection tool highlights the image that appears to be changed, and it can go further. As the training is based on a well-established Adobe tool, artificial intelligence can reverse the changes to produce a very good approximation of the original photo.
This is an impressive first step in the fight against misinformation on the Internet, but there are countless ways to create fake photos. This tool will not detect anything other than the Photoshop liquefaction filter. It will take a lot of work to create an AI capable of taking into account all or more image manipulation techniques. This is probably a good way to research, since humans are so bad at detecting fake ones.