In recent years, technology has had a large impact on the fight against human trafficking and efforts to prosecute those responsible for the sale of minors. Online portals frequented by traffickers have seen an increase in the number of images of victims of human trafficking.
A major issue in efforts to track down both criminals and victims is that analysts are not able to determine where an image of the situation was taken. This is made more difficult considering that images are taken in rooms with little to no distinct characteristics such as bare hotel rooms or basements. Therefore, police and investigators are unable to determine the location of rooms used for such criminal activity. However, using artificial intelligence, a team of researchers from George Washington University, Temple University, and Adobe Research have found a way to identify rooms and locations to combat human trafficking. The team developed a database, Hotels-50K, that compiles over 1 million hotel images from over 50,000 different hotels around the world including the United States, Western Europe, and coastal regions. Images in the Hotels-50K data set are compiled from travel websites and crowd-sourced images from mobile applications which can then be compared to photos from real-world investigations.
There are challenges that come with trying to identify a hotel room from a single image due to poor image quality, poor camera angles, large occlusions, and indistinguishable objects such as furniture, art, and bedding. But within the Hotels-50K database, artificial Intelligence, large image data sets, and scene and place recognition are all utilized to combat this issue. The Hotels-50K database includes specific metadata such as hotel name, geographic location, hotel chain, or “Other” when the hotel is not part of a major chain. And to further strengthen the usefulness of Hotel-50K, the database incorporates images from TraffickCam, a database-like app that allows users to upload images of hotel rooms for investigators to efficiently search images.
To test the efficiency of the platform in identifying case photos based on the Hotels-50K data set, researchers have created a set of test images to replicate images from real-world investigations as closely as possible. The test set consists of roughly 17,954 images from over 5,000 different hotels on TraffickCam. In real world investigations before analysis, investigators crop victims out of photos to protect victims’ privacy, a process known as ‘masking’. The researchers performed ‘masking’ at four different levels: none, low, medium, and high. In highly masked photos, the cropped area of the photo can take up 85% of the image making it much harder to identify. From later testing using two pre-trained neural network data sets, it was concluded that Hotels-50K identifies common hotel chains with an 80% accuracy rate. But while the data set was able to identify nearly one thousand images and their associated hotel chain, new challenges appeared when trying to identify the specific hotel. The data set was only able to identify the first 100 images 24% of the time when attempting to locate the specific hotel. While 24% falls well below what one would imagine as an ideal success rate, these chances already far surpass previous image identification capabilities.
The baseline approach of Hotels-50K is now being utilized by The National Center for Missing and Exploited Children to combat human trafficking, and novel algorithms may be utilized to enhance search performance in ongoing investigations.