Algorithms that can quickly match images or even help identify their origin are being deployed to track down criminals who trade in child pornography
CHARLES M. JOHNSON JR is today in a US federal jail for producing child pornography. His conviction came thanks to a sharp-eyed investigator at the (NCMEC) in Arlington, Virginia, who spotted that two images among the millions received by the centre each year showed different pre-pubescent girls being sexually assaulted on the same distinctive bedspread.
“We identified the manufacturer of the bedspread, and where it was sold,” recalls , NCMEC’s president. Zeroing in on Indianapolis, Indiana, and working with the local police, the children were identified. Johnson was a live-in babysitter with one of the girls’ families, and had moved with them when they relocated to Cincinnati, Ohio, where he was tracked down and apprehended. When Johnson was arrested in December 2005, thousands of pornographic images of children were found on his computer.
Advertisement
Now, thanks to software developed by Google, NCMEC no longer has to rely on feats of memory from its small team of investigators to make similar connections. “Google developed a tool that they call the ‘bedspread detector’,” says Allen. Using advanced pattern-recognition algorithms, the software can search NCMEC’s database of images, matching those that contain common background features such as potted plants, framed pictures – or bedspreads.
This is just one way in which technology companies are helping NCMEC and law enforcement agencies combat the growing problem of child pornography, a lucrative business that is controlled in part by organised crime syndicates. “The internet, unfortunately, has allowed for the ‘mainstreaming’ of child pornography,” says , a specialist in media law at the University of Florida in Gainesville, who has . “The question is now what technology can do to put away the individuals who traffic this kind of content.”
NCMEC is at the heart of this effort. Although it is a private body, the centre has been designated by Congress as the national clearing house for information on exploited children. In 2002 it established the , which receives photos and videos from child pornography investigations across the US, and attempts to identify the children involved.
Over the past eight years, the volume of material NCMEC has received has risen inexorably (see graph), and statistics on cases coming before federal courts reveal a similar upwards trend (see graph). Without the help of the technology firms that are part of NCMEC’s , such as Google, Microsoft and Yahoo, Allen admits that “we would absolutely have been overwhelmed”.
Google’s involvement ramped up when a team led by visited NCMEC in 2007. “I thought I’d seen it all before,” says Baluja, who had previously worked on software to filter out pornographic material from search results. “But I was absolutely horrified by the type of stuff they have to deal with.”
Spurred into action, Baluja’s team adapted Google’s visual search algorithms to suit NCMEC’s needs. In addition to the bedspread detector, the tools Google donated to the centre the following year included the code that underpins a system known as , originally developed to detect violations of copyright on . This allows NCMEC to identify video sequences in newly received material that match those already in its database.
It’s not just Google that is helping the fight against child pornography. Microsoft, too, has developed tools to help track down offenders. One such piece of software is the (CETS), which was unveiled in 2005. This is a database tool that allows law enforcement agencies to share evidence from their investigations, allowing vital connections to be made. Now in use in countries including Australia, Brazil, Italy and the UK, CETS has helped to secure dozens of arrests.
More recently, Microsoft teamed up with , a specialist in digital imagery at Dartmouth College in Hanover, New Hampshire, to develop a system called . This can scan through millions of images being stored and viewed online, flagging those that match images held in NCMEC’s database. That’s useful because child pornographers repeatedly circulate old images as well as new ones. The idea is that PhotoDNA will help companies that unknowingly host illegal material to detect and purge it – and pass information to the police for investigation.
Computational simplicity is crucial, so that vast numbers of images can be scanned without slowing online networks. As a result, PhotoDNA doesn’t search visually. Instead, it matches images based on digital signatures – strings of numbers created by an algorithm from the underlying data in the image file.
But conventional digital fingerprinting systems create signatures that are too specific to be useful – resize or compress an image, or even alter a single pixel, and it has a different signature and can’t be matched to the original. To get round this problem, Farid and Microsoft used an algorithm that creates “fuzzy” signatures that can be used to reliably match images depicting the same digital photograph, even if they have been subtly altered.
Over the past year, Microsoft has been putting PhotoDNA through its paces by using it to match a relatively small sample of NCMEC’s images against the images in web pages indexed by the search engine, and material on , which includes the online data storage service. According to Microsoft, PhotoDNA is identifying illegal images with a false-positive rate of less than 1 in 1.5 billion. Facebook has already licensed the technology for free from NCMEC, says Allen, and other companies have asked to evaluate it.
Although Microsoft has not yet released the results of its PhotoDNA trials, Farid says that those involved were shocked by the volume of child pornography identified. “Most people think this is relegated to a dark, tiny corner of the internet, but the scope of the problem is much larger,” he says.
“Most people think child pornography is relegated to a tiny corner of the internet, but it isn’t”
Turning the tide may require even more cooperation from industry. In testimony before the US Congress in January, Jason Weinstein of the US Department of Justice that investigations frequently go cold because internet service providers quickly delete data on their customers’ online activities.
Allen sees the fight as an arms race, in which NCMEC and its allies must innovate to keep pace with the enemy. A top priority is identifying images and video being shared on peer-to-peer networks, which are inherently harder to probe than material held on a central server. “It’s so important that we stay on top of these technologies,” he says.
Should virtual child porn be illegal?
People are good at telling the difference between computer-generated images and real photographs, according to an unpublished study by Hany Farid of Dartmouth College in Hanover, New Hampshire. Volunteers asked to view colour photographs and CGI correctly identified the latter between 85 and 95 per cent of the time, for images larger than 50 pixels across.
The research may have implications for prosecuting child pornography cases in the US. In April 2002, the US Supreme Court wording from a that had attempted to outlaw “virtual” child pornography. The court ruled that pornographic CGI involving no actual children was protected free speech.
In response to the ruling, some defendants withdrew guilty pleas, hoping to argue a “virtual” defence. It had little effect – most suspects are caught in possession of images and video in which some of the children involved can be identified.
But the issue could return. Following the Supreme Court decision, Congress enacted a , defining child pornography as including any images that are “indistinguishable from” real photographs.
As Farid’s research shows, CGI technology is not yet at the point that most people are fooled into thinking that CGI images are real photos. But it is constantly improving. “Five years from now we’re going to have to redo the study,” says Farid.
When CGI can match reality, the issue of whether virtual child pornography should be treated the same as real images of abused children may once again be challenged.
