http://www.digitaljournal.com/tech-and-science/technology/adobe-detects-manipulated-images-using-ai/article/525591

Adobe detects manipulated images using AI

Posted Jun 26, 2018 by Tim Sandle
Adobe, the company that brought the world Photoshop, has developed an artificial intelligence which can tell whether an image has been manipulated. The platform can identify if an element had been added, moved or cut from an image.
Using the Adobe Photoshop Touch application on an Apple iPad
Using the Adobe Photoshop Touch application on an Apple iPad
Adobe Systems
The new platform has been developed by Vlad Morarium, a researcher at Adobe researcher. Morarium has demonstrated how artificial intelligence can be used to to scan for signs of manipulation, on a scale that is not visible to the unaided human eye.
[url=https://theblog.adobe.com/spotting-image-manipulation-ai?origref=https%3A%2F%2Fwww.bbc.co.uk%2Fnews%2Ftechnology-44601469 t=_blank]Writing on the Adobe blog, Morarium says: "We focused on three common tampering techniques—splicing, where parts of two different images are combined; copy-move, where objects in a photograph are moved or cloned from one place to another; and removal, where an object is removed from a photograph, and filled-in."
The basis of the research is that each image manipulation technique, of which the common ones, according to the BBC, are: "splicing, where parts of two different images are combined; copy-move, where objects in a photograph are moved or cloned from one place to another; and removal, where an object is removed from a photograph, and filled in", leaves an artifact. The developed algorithm has been trained to detect these artifacts.
He explains more in the following video:
As the video indicates, the techniques used in the new research provide more possibility and new options for managing the impact of digital manipulation. For those concerned about the accuracy of images, as with many in the media, the approach could potentially answer questions of authenticity more effectively.
The system used is outlined in a white paper. Here Morarium and colleagues describe how they used a two-stream Faster R-CNN network and trained it end-to-end to detect the tampered regions given a manipulated image. One of the two streams is an RGB stream, which works to extract features from the RGB image input to find tampering artifacts. The second stream is a noise
stream that leverages the noise features extracted from a model filter layer to discover the noise inconsistency. The paper is titled "Learning Rich Features for Image Manipulation Detection."