The ISIS recruits supporters and fellow travelers from around the world largely by spreading photos and videos of its violent exploits online. What if social-media companies could automatically detect and delete such imagery?
The Counter Extremism Project, working with Dartmouth University computer scientist Hany Farid and funding from Microsoft, have developed a new method for doing just that. They hope to provide the software to help companies like Twitter, Facebook, and Google stop extremist groups from distributing such material on social media.
Based on a concept called robust hashing, the idea was invented in 2008 by Farid, who was trying to help stop the flow of child pornography. But telling a machine to recognize a picture of child abuse “is not possible,” he told reporters on a conference call on Friday. “We have not gotten to the stage where we can reason about fairly high-level things having to do with content. And so they were stuck.”
Consider that Google and Stanford needed a neural network of some 16,000 processors looking at more than 10 million YouTube videos before the system could recognize a cat.
“Deep learning, and other related modern-day learning algorithms are not capable of quickly and reliably distinguishing between appropriate and inappropriate content,” Farid told Defense One in an email. “The speed and error rate for these approaches is simply prohibitive for dealing with Internet-scale image uploads” — billions of new images per day.
So rather than teach a network of machines to understand the concept of abuse, Farid sought to teach software to recognize images that were spreading across the Internet after someone had tagged them as inappropriate. This way, the computer must simply recognize the image, not the idea that the image is depicting, and stop it from proliferating further.
Farid describes robust hashing as a method of extracting a “distinct signature” from an image. The signature is “an abstract numerical representation that is both distinct and stable to modifications to the underlying medium,” meaning that it remains useful even if the image is modified a bit. He further described the signature as “a string of numbers that embody that actual, underlying content,” like a photo’s DNA.
Farid donated the software to the National Center for Missing & Exploited Children, which licenses it for free to social media companies to police their sites for pictures of child pornography. It has an error rate of about one in five billion, he says.
About eight months ago, Farid teamed up with the Counter Extremism Project to apply robust hashing to a different challenge: recognizing terrorist content. They trained a system on the Project’s ever-expanding propaganda-images database, and set it to work.
“The technology will be running on any company that chooses to participate in this process. Associated with the technology is a database of hashes of known content. That hash table can be distributed easily through any number of different means and so every time there is new content, the hash table gets expanded, that hash table gets distributed to the companies that are participating and then the process is fully automatic at that point,” he explained, Defenseone reported.
212-11