By DAVID BROOKS
Monitor staff
When Hany Farid started working a few decades ago on analyzing digital images, he was thinking about how to detect forgeries and copies of still photos and probably hadn’t even heard of the word “jihad.”
Fast forward to 2016. The Dartmouth professor’s research has now morphed into something called robust hashing (“hash” being a computer term roughly similar to a digital file’s signature), and is being touted as an important new tool against online videos of violence, beheadings and anti-Western diatribes that are used to recruit terrorists and spread violent hate.
Such videos are embraced by groups like the Islamic State as an easy way to radicalize people and get them to join the cause. The goal, then, is to remove them before they can be spread around social media sites.
Easier said than done.
“As soon as you take one of these videos down, they get posted in 20 other places. You can play whack-a-mole, but you’ll never win,” said Farid, who heads the computer science department at Dartmouth.
His solution is a software tool that creates a unique fingerprint for any video that can be used to automatically find copies as soon as they’re uploaded, without waiting for a person’s judgment on similarity.
This may sound familiar, since Farid previously developed a system called
PhotoDNA used to automatically identify child pornography pictures online, but he says the technology is entirely different, partly because of the scale involved.
“The trick here is that unlike a single image, for every second of video you have 24 still images; you have a massive amount of data. When I watch Netflix, I’m amazed that it works. When I think of the sheer number of pixels that have to be sent, it’s insane that it works,” he said.
Further, video is also more malleable than still images.
“It’s tricky. Videos get cut and spliced as they move around the internet. A video shows a beheading, they’ll just cut out a little part of it, maybe add a teaser, something at the end – the video is very different, not even the same length, but you still have to identify it,” he said. “That’s the technology: The ability not just to match videos in their entirely but just little chunks of video. And the same thing is true of audio.”
“We need to extract a digital signature that is stable over the lifetime of the video, when it has been recompressed, spliced, with additions, subtractions, that signature stays constant; that is distinct, so a video of kittens will not be confused with video of beheading; and is efficient, that can run at internet scale,” he said. “Even a 0.1 percent false alarm rate is an astonishing rate of failure at internet scale.”
Farid says he’s been working on the process for a year or more on the side of his job as a professor – in which, among other things, he teaches CS1, the intro computer science class ,in front of a few hundred students. Not exactly glamorous work, but he says he loves it.
Farid came to Dartmouth two decades ago, and I first talked to him in 2003 (yes, that long ago – the realization made us both feel a little old) when he had drawn attention for groundbreaking work creating the field of digital forensics, with a goal of detecting manipulation of digital pictures. He later established a private company called Fourandsix (a play on the word “forensics”) to commercialize the technology. He says it hasn’t exactly set the world on fire, but it’s still around.
Farid says he turned to the issue of child pornography in 2008 after Microsoft approached him.
“I got a call from folks at Digital Crimes Unit at Microsoft. I remember thinking at the time I don’t have time for this, but it turned out to be one of the single most important things I’ve done in my career,” he said.
In recent weeks, Farid’s video work has drawn a flurry of attention following news it was being adopted by the Counter Extremism Project, a policy group designed to fight extremism, with a financial boost from Microsoft.
It’s a two-step process. Videos have to be tagged as related to terrorism by a person. Then a robust hash is created and placed in a database, overseen by the Counter Extremism Project, that can be used by the Facebooks and Twitters of the world to keep copies from appearing.
There is understandable unease about using software to automatically remove videos, both for technical reasons – it’s easy to imagine how an instructional video about harvesting pumpkins could be tagged as showing a beheading – and for social reasons, because one person’s terrorism might be another person’s social protest.
Farid says the robust hashing will prevent the technical problem, and as for the social concerns, they already exist.
“There are gray areas – we do not want to quash dissent or public discourse, just beheadings, torture, explicit calls to violence,” he said.
Farid noted that many online and social media services already take down these videos because they violate terms of service. But they do it manually, meaning they don’t do it well. He hopes to improve that.
“As an academic, it’s rare you can do something (that) within a relatively short amount of time, you see it have an impact on this kind of global internet scale,” he said. “It’s very satisfying.”
(David Brooks can be reached at 369-3313, dbrooks@cmonitor.com or on Twitter @GraniteGeek.)