One photo of a child in a hotel room can often be the only clue to a trafficked child鈥檚 whereabouts. An artificial intelligence is now helping investigators to identify these hotel rooms, leading to the rescue of a number of sexually exploited children.
Globally, an estimated 4.8听million people have been forced into sexual exploitation. More than 1 million are under 18.
In the US, exploited children often appear pictured in hotel rooms in online adverts. These images are found across dozens of听websites as well as on dating apps. Traffickers regularly move听location to try to avoid being found.
Advertisement
To fight back, Abby Stylianou at听George Washington University in Washington DC and colleagues built an AI that attempts to identify hotels from these adverts. It does this by comparing the advert images to a听database of more than 1 million photos of 50,000 hotels around the world, including some from travel websites and others sent in听by volunteers.
To train the AI, the team adjusted some of the images to resemble trafficking photographs by cropping, rotating and altering the colour. They then blacked out听parts of the images with silhouettes to resemble a person in the foreground.
The images donated by volunteers were particularly useful, says Stylianou. That is because they have similar lighting to those taken by children who are听coerced into taking photos of听themselves.
Regular renovations and hotel chains with identical decor makes the task more difficult, as does the听fact that many images the investigators find have much of听the background obscured.
Like finding a needle in a haystack
In tests on images the AI hadn鈥檛 previously seen, it identified the correct hotel chain 63 per cent of听the time in a top-five list by similarity. However, identifying the specific hotel was more difficult. When producing a list of听100 candidates, the AI only included the correct hotel around 25 per cent of the time.
While the accuracy of the system still needs improving, trafficking investigations are like听finding a needle in a haystack, says Staca Shehan at the National Center for Missing and Exploited Children (NCMEC) in the US. Any听tool that can narrow leads of听enquiry down is extremely useful, she says.
Over the past year, NCMEC has successfully used the AI to find and rescue an undisclosed number of exploited children.
Before the AI, NCMEC investigators would try to find similar images using Google, but this typically doesn鈥檛 find images from the correct hotel.
鈥淚f that failed, they鈥檇 then try to manually search through images on travel websites if they know the city or region of interest,鈥� says Stylianou. This is a painstaking and time-consuming process. After that, there may be appeals to听the public, but successfully identifying a hotel is contingent upon someone seeing and recognising the location.
Another use for the AI is for building evidence in court cases against human traffickers.
If someone is trafficked between two or more US states it听is a federal, rather than state, crime and warrants harsher penalties. Courts often rely on听a听child鈥檚 testimony of travel patterns and exploitation over time, and results from the AI听could be used to support that听testimony.
Tools like this are a valuable public resource, says Chris Mattmann, a data scientist at NASA, who has worked on similar tools. He says that others can build on the system to help make听it more accurate.
Automatic Investigator
An AI that can identify hotel rooms (see main story) isn鈥檛 the only way machines are helping investigators.
Last year, one team unveiled an AI that could track down potential child abusers in large organisations by looking for suspicious patterns. It is inspired by a team at The Boston Globe who used a similar approach to听identify child abuse in the Catholic church, which was made famous by the film Spotlight.
The system analyses career trajectories from decades of documents and tries to identify outliers that could indicate a cover-up, such as people who suddenly go on sick leave or move around a lot. This is then investigated听by humans.
Another AI inspects the wording in听online sex adverts to try to find common authors. The idea is that people who post multiple adverts have a higher likelihood of being traffickers rather than sex workers.
When tested on 10,000 adverts, it correctly identified about 90 per cent of those posted by the same author.
Reference:听arXiv,听
