Blink twice if you are a fake. Sophisticated fake videos created by artificial intelligence are getting better and better, so tools to identify the genuine article are needed. One solution is to look at the way fake faces blink.
In recent months, eerily-realistic fake videos have been created of Theresa May, Barack Obama, Vladimir Putin, and Donald Trump giving speeches, as well as of famous actors superimposed into porn scenes. To make the footage, AI needs still images of the faked person, which are then mapped onto an actor playing out the scene. The results are often difficult to distinguish from the real deal.
However, all is not lost. There may be a simple tell for computerised fakery – the people in deepfakes videos don’t blink like humans.
Advertisement
The average person blinks 17 times a minute, but because most photographs don’t capture people with their eyes closed, the subjects of deepfakes videos don’t close their eyes either. “The deepfakes algorithm inherits this basic flaw,” says Siwei Lyu, at SUNY Albany.
So Lyu and his colleagues created a new fakery fighting tool that tracks the eyes in videos. Using this approach, the team’s tool managed to achieve an accuracy of 99 per cent.
Though the tool works for now, it may not remain that way for long, says Eliot Higgins, at the Atlantic Council’s Digital Forensic Research Laboratory. “If they can make someone say something they’re not saying, I think making someone blink isn’t going to be particularly difficult,” he says.
Stanford University researchers have already developed a better version of deepfakes that includes blinking and will be presented at the ACM SIGGRAPH, a computer graphics in August. That doesn’t dissuade Lyu though. “Media forensics is a cat and mouse game,” he says.
Journal Reference:
