Advances in artificial intelligence have made it increasingly easy to create visual counterfeiting, both in image and video form. Now, however, there is a new AI technology that will help combat precisely this.
In its official blog, Adobe now announces that in collaboration with researchers at the University of Berkeley, they have developed a system that can detect the manipulation of digital images using artificial intelligence.
Manipulated by Photoshop on your own
The researchers' method was to train the so-called "Convolutional Neural Network," which is a complex type of neural network inspired by the behavior of the brain that deals with the processing of visual impressions.
The system was trained with a database consisting of a large number of Internet images that Photoshop has been configured to automatically manipulate through the program's own "Face Aware Liquify" feature – a feature used to change face features.
In addition to this he was also a professional digital artist hired to manipulate arbitrary images contained in the database. The point of also making use of human creativity was to achieve greater breadth in the manipulation of images.
To test the system researchers wanted to compare it with the level of accuracy for people. This was done by displaying the pair of images of people who knew one of the images was manipulated digitally. According to the researchers, test subjects were able to detect digital manipulation in 53% of cases, which is only slightly better than pure divination.
achieved accuracy level of 99 percent
The new AI system was, however, able to achieve results of as much as 99 percent through various experiments. In addition, the tool was able to identify the specific methods of changing the faces and which areas of the face were manipulated.
The system was able to perform a "repeating function" by reversing the images back to their original state – with a level of accuracy that the researchers themselves did not expect.
While much work still needs to be done, the researchers behind the new project believe the tool represents a breakthrough to allow people to verify the authenticity of digital expressions.
– This is an important step to be able to detect certain types of image editing and undo fungererer surprisingly well. In addition to such technologies, the best defense is a sophisticated audience who knows that content can be manipulated – sometimes to please, but sometimes to cheat, said Adobe's research department head Gavin Miller.
All the technical details of how the system works can be found in the search document, and also see the video presentation below.
Read also: When will AI be beneficial intelligence? »