Online social networks (OSNs) represent powerful digital tools to communicate and quickly disseminate information in a non-official way. As they are freely accessible and easy to use, criminals abuse of them for achieving their purposes, for example, by spreading propaganda and radicalising people. Unfortunately, due to their vast usage, it is not always trivial to identify criminals using them unlawfully. Machine learning techniques have shown benefits in problem solving belonging to different application domains, when, due to the huge dimension in terms of data and variables to consider, it is not feasible their manual assessment. However, since the OSNs domain is relatively young, a variety of issues related to data availability makes it difficult to apply and immediately benefit from such techniques, in supporting the detection of criminals on OSNs. In this perspective, this paper wants to share the experience conducted in using a public dataset containing information related to criminals in order to both (i) extract specific features and to build a model for the detection of terrorists on Facebook social network, and (ii) to highlight the current limits. The research methodology as well as the gathered results are fully presented and then the data-related issues, emerged from this experience, are discussed.
.