A fake is only as good as it looks. But while forging a counterfeit handbag or watch takes time and effort, churning out fake videos has become surprisingly easy.
A new system can turn a few simple animated line drawings into realistic fake clips in high definition. The software is open source, meaning that it is available to anyone 鈥� and it has reignited concerns that such tools could be used to warp our perception of the world.
The new system was created by researchers at graphics hardware firm Nvidia and the Massachusetts Institute of Technology. It works by exploiting the special features of a generative adversarial network, which is a type of artificial intelligence.
Advertisement
First, the system must be trained with hundreds of videos of a certain type of scene, such as a driver鈥檚 view of a journey along a street or a person talking to a camera. It then breaks a real video down into component layers, such as foreground and background, and key elements, like buildings and trees. This provides a basic, editable structure onto which fake elements can be added.
Then, one part of the AI generates new frames of a new video sequence聽based on an input video such as basic animated sketches. Another attempts to distinguish the generated frames from real video frames. If it can鈥檛, those new frames are judged to be good enough and are inserted into the synthesised video.
The resulting footage can be produced at 2K resolution and looks startlingly lifelike. Examples the team has produced include , and .
鈥淚t is sort of stunning, the progress that has been made,鈥� says digital forensics expert at Dartmouth College.
This type of video has become known as a deepfake. Recently, fake videos of world leaders, such as Donald Trump and Theresa May, were created using similar techniques. A community dedicated to creating fake porn videos containing famous actors has sprung up too.
Until now, the technology to create deepfakes had largely been private or had produced cartoonish results, says artist and programmer . That has now changed.
鈥淚t鈥檚 possible that people might use this to create content that isn鈥檛 actually real and could be used to mislead people,鈥� says Bryan Catanzaro at Nvidia. But he explains that his team鈥檚 interest in developing the tool was rather less nefarious. They wanted to find out if AI could render realistic video game graphics from lifelike sources. Street scene video footage of world cities could be used to build levels in a driving game, for instance.
However, will this latest system be used to convince us that a politician said something they didn鈥檛, for instance? Farid says distinguishing real videos from fake ones is now a 鈥渟ignificant problem鈥�. However, he also thinks that informed fact-checkers will likely still be able to detect fakes for some time to come.
How to spot a fake
There are also some tell-tell signs that fakes exhibit. One research team recently pointed out that faked videos of politicians could be detected by checking whether the individual blinks at a realistic rate. The new tool also still creates videos that occasionally contain artefacts and giveaways, including continuity errors 鈥� sometimes a car driving along a road will change colour gradually.
The rise of fake videos could also have a counter-intuitive effect: knowing such videos exist, people will more comfortably discount things that are in fact real 鈥� like when negative headlines are pooh-poohed as fake news. 鈥淭hat in some ways is the real threat,鈥� says Farid, 鈥渢hat we don鈥檛 believe anything anymore.鈥�
One example of this is in the US where people caught with videos containing the sexual abuse of children聽have claimed 鈥� known as the virtual defence. Holding such computer generated聽videos is still a crime, but comes with a shorter sentence.
However, there are many legitimate cases for technology like this, says Kogan. For example, the same principles can be used for synthesising architectural designs after supplying the system with a basic sketch of a building, which could be helpful for designers.
Gaurav Oberoi at the Allen Institute for AI that celebrities could 鈥渓icense their faces鈥� for use in advertising, without having to actually be present for filming.
Journal Reference:聽
