Deep Learning Helps Fight the Spread of Child Pornography

Most social media websites and apps have some sort of automated system to identify and remove offensive, inappropriate, and illegal images. Instagram, for example, has 60 million new pictures uploaded every single day, and it would be impossible to have humans manually check each one. But, the system that they use is proprietary, so Christian Haschek set out to build his own detection service.

Haschek created the open source image hosting service PictShare, and was troubled when it was brought to his attention that someone had uploaded child pornography to the service. He rightfully contacted the police, who gave him the very dubious advice to print the image and take it to his local police station — an act which could be illegal in its own right. Realizing that, he instead contacted INTERPOL and was told to delete the offending image and report the IP address of the uploader. He then came to the conclusion that there could be more images like this on his service, so created the NSFW as a Service system to find them.

The system uses a deep learning neural network that was trained on the Open NSFW model created by Yahoo to detect nudity. It runs on a Raspberry Pi in conjunction with an Intel Movidius neural compute stick, which helps offload the CPU-intensive task of running a neural network. When it runs, it builds a queue of images that have more than a 30% certainty of containing nudity, which can then be checked by a person and reported to the authorities. When NSFW as a Service was used on another major image host, 3,292 child pornography images were found. It is, of course, horrible that these images exist in the first place, but Haschek’s service is very important for limiting their spread, which in turn should make it less likely that they’ll ever be created.