Learning from photos, a deep neural network identifies deepfakes
They're known as deepfakes – photos or videos that have been very convincingly manipulated to depict people saying or doing things that they never actually said or did. They're potentially quite the problem, so an experimental new deep neural network has been designed to spot them.
Led by Prof. Amit K. Roy-Chowdhury, a team at the University of California, Riverside started with a large dataset of both manipulated and non-manipulated photos. The researchers already knew which ones were which, and computer-labelled them accordingly.
In the manipulated pictures, they highlighted the pixels along the boundaries of objects that had been digitally added to the shot – it had previously been established that in faked photos, those boundaries tend to be smoother or otherwise different than those of objects that were actually in the shot when it was taken. Although those differences can't necessarily be detected by the human eye, a pixel-by-pixel examination done by a computer will pick them up.
The labelled dataset was then fed into a deep neural network, which is a set of algorithms modelled loosely after the human brain, designed to recognize patterns in raw data. Using the images, that network learned to identify the telltale boundaries of digitally-added images. When it was subsequently shown photos from outside of the dataset, that it hadn't seen before, it was able to spot the fakes "most of the time."
Although the system currently only works on photos, the team is now working on applying it to videos, where it would most likely just analyze individual still frames. That said, the technology may never be 100-percent accurate, and could likely end up being used to flag suspicious images that are subsequently analyzed by people.
"If you want to look at everything that's on the internet, a human can't do it on the one hand, and an automated system probably can't do it reliably," says Roy-Chowdhury. "So it has to be a mix of the two."
A paper on the research was recently published in the journal IEEE Transactions on Image Processing.