Technical report on the Automatic Detection of Paedophile Queries

Filtering or identifying paedophile queries is a key issue for law enforcement and search engines. However, these queries are in general mixed with a huge amount of other queries. Moreover, little is known on their characteristics. We address here these two issues in order to design the first tool for automatic detection of paedophile queries. Using domain expertise, we select some paedophile queries in a set of hundreds of millions of queries entered in a general public P2P system. We extend this set by manually inspecting the queries it contains and the words composing them. We then design a tool which tags any query as paedophile or not. We run it on our dataset and evaluate its performances by submitting appropriate samples to external experts. This assessment shows that the tool performs very well. Going further, the assessment makes it possible to estimate precisely and rigorously its error rates, which we compute and provide.