Adobe, the company which brought the analog photographs into the digital world, 28 years ago and changed everything. Adobe developed its most popular photo editor, Adobe Photoshop, a synonym for photo editors, for editing photos like never before. And it is still being used to edit and create the artistic expression in reality. And as everything has its own pros and cons, Photoshop is being used to edit and create fake content, and sometimes the doctored content may worsen the situation. But now Adobe is working to make it easier to find out the photoshopped content easily with the help of Artificial Intelligence (AI).
Adobe showcased its latest research work at the CVPR computer vision conference this month. The research papers published do not show a sudden discovery, as the research is in its early phase, and describes how the Artificial Intelligence can fasten the process of digital forensics, which is being done by the humans. Adobe says AI is the new solution to the old problem.
It’s the first time that Adobe is researching to create a software for Digital Forensics, but normally there is human involvement needed for this type of tasks. Vlad Morariu, senior research scientist at Adobe, who has been working on technologies related to computer vision for many years, explains that a number of tools exist that can be used to find out the manipulation or to find out if the content is photoshopped or not. He says, that the file formats contain metadata that can be used to store information about how the image was captured and manipulated. There exists some forensic tools which are used to detect the manipulation, just by examining the digital artifacts, like inconsistency in the color palette, or variation in color and brightness, strong edges, lighting and the pixel value of the photograph. For example, if cloning ( copying a part of the image and paste it in another part) is done on a photograph, then experts look for the traces like inconsistency in the pattern, and if layer cropping is done, experts look for the variation in the brightness in the photos. But it’s not easy, as it looks.
To find out any manipulation in the photograph, there is a need for both digital forensic tool and expert. But even having the best combination of both, there is a probability that the manipulation in the image can’t be found out, as the tools do not provide the deep understanding of the photo, and the tool can’t be used for every situation. Some tools are useless whereas, some need deep expertise and a lengthy execution. So, technologies like Artificial Intelligence and machine learning can be used to detect if the image has been modified or not, and it can be done not only easily but quickly and effectively.
When an image is modified or manipulated, it tends to leave some traces that can be studied using tools and by this, it can be studied, how the image is altered. Experts search for smoothed areas, strong contrast edges and difference in noise patterns. These type of traces can’t be observed by human eyes, as these needs to be observed at the pixel level.
Adobe focused on three common tampering techniques which are done to photoshop an image – first is, Splicing, where parts of two or more different photographs are combined, second is Copy-Move, where object or any part of the image is moved or cloning where a part of the image is copied and paste in the same image and Removal, where an object or the part of the image is removed from a photograph, and any part of the image is filled-in.
Adobe is using Artificial Intelligence and machine learning to detect the traces on the pixel level. For this company has created a large dataset of tens of thousands of examples of doctored images and their pattern of alteration to train the deep learning neural network to recognize image manipulation. The research is successful so far, it’s not a breakthrough. But, for now, the deep learning neural network has no direct application to detect manipulation in the deepflakes, or we can say the edited video created using AI.
The digital forensics experts say that the AI technique works well to detect the manipulation but it’s not ready for all the conditions for now, as still there is a lot that the system has to learn, or we can say a lot more data needs to be fed in the neural network to learn high-level manipulation done to the photograph.
The overall new research by Adobe still needs a lot of improvement to be used in place of the tools and experts completely in the digital forensics.