Seeing shouldn’t always be believing as faked videos become more realistic
LOS ANGELES—All it takes is a single selfie.
From that static image, an algorithm can quickly create a moving, lifelike avatar: a video not recorded, but fabricated from whole cloth by software.
With more time, Pinscreen, the Los Angeles startup behind the technology, believes its renderings will become so accurate they will defy reality.
“You won’t be able to tell,” said Hao Li, a leading researcher on computervideo at the University of Southern California who founded
Pinscreen in 2015. “With further deep-learning advancements, especially on mobile devices, we’ll be able to produce completely photoreal avatars in real time.”
The technology is a triumph of computer science that highlights the gains researchers have made in deep neural networks, complex algorithms that loosely mimic the thinking of the human brain.
Similar breakthroughs in artificial intelligence allowed University of Washington researchers to move President Barack Obama’s mouth to match a made-up script and the chipmaker Nvidia to train computers to imagine what roads would look like in different weather.
What used to take a sophisticated Hollywood production company weeks could soon be accomplished in seconds by anyone with a smartphone.
Not available for a video chat? Use your life-like avatar as a stand-in. Want to insert yourself into a virtual reality game? Upload your picture and have the game render your character.
Those are the benign applications.
Now imagine a phony video of North Korean dictator Kim Jong Un announcing a missile strike. The White House would have mere minutes to determine whether the clip was genuine and whether it warranted a retaliatory strike.
What about video of a presidential candidate admitting to taking foreign cash? Even if proved fake, the damage could prove irreversible.
In some corners of the internet, people are using open-source software to swap celebrities’ faces into pornographic videos, a phenomenon called Deep Fakes.
It’s not hard to imagine a world in which social media is awash with doctored videos targeting ordinary people to exact revenge, extort or to simply troll.
In that scenario, where Twitter and Facebook are algorithmically flooded with hoaxes, no one could fully believe what they see. Truth, already diminished by Russia’s misinformation campaign and President Donald Trump’s proclivity to label uncomplimentary journalism “fake news,” would be more subjective than ever.
The consequences could be devastating for the notion of evidentiary video, long considered the paradigm of proof given the sophistication required to manipulate it.
“This goes far beyond ‘fake news’ because you are dealing with a medium, video, that we traditionally put a tremendous amount of weight on and trust in,” said David Ryan Polgar, a writer and self-described tech ethicist. “If you look back at what can now be considered the first viral video, it was the witnessing of Rodney King being assaulted that dramatically impacted public opinion. A video is visceral. It is also a medium that seems objective.”
To stop the spread of fake videos, Facebook, Google and Twitter would need to show they can make good on recent promises to police their platforms.
Last week’s indictment of more than a dozen Russian operatives and three Russian companies showed how easily bad actors can exploit the tech companies that dominate our access to information. Silicon Valley was blindsided by the spread of trolls, bots and propaganda—a problem that persists today.
Tech companies have a financial incentive to promote sensational content. And as platforms rather than media companies, they’ve fiercely defended their right to shirk editorial judgment.
Critics question whether Facebook, Google and Twitter are prepared to detect an onslaught of new technology like machine-generated video.
“Platforms are starting to take 2016-style misinformation seriously at some levels,” said Aviv Ovadya, chief technologist at the Center for Social Media Responsibility. “But doing things that scale is much harder.”
Fake video “will need be addressed at a deeper technical infrastructure layer, which is a whole different type of ballgame,” Ovadya said.
(Facebook and Twitter did not respond to interview requests. Google declined to comment.)
The problem is that there isn’t much in the way of safeguards.
Hany Farid, a digital forensics expert at Dartmouth College who often consults for law enforcement, said watching for blood flow in the face can sometimes determine whether video is real. Slight imperfections on a pixel level can also reveal whether it is genuine.
Over time, though, Farid said, artificial intelligence will undermine those clues, perpetuating a cat-and-mouse game between algorithms and investigators.
“I’ve been working in this space for two decades and have known about the issue of manipulated video, but it’s never risen to the level where everyone panics,” Farid said. “But this machine-learning-video has come out of nowhere and has taken a lot of us by surprise.”