A Semi-supervised Approach to Measuring User Privacy in Online Social Networks

During our digital social life, we share terabytes of information that can potentially reveal private facts and personality traits to unexpected strangers. Despite the research efforts aiming at providing efficient solutions for the anonymization of huge databases (including networked data), in online social networks the most powerful privacy protection is in the hands of the users. However, most users are not aware of the risks derived by the indiscriminate disclosure of their personal data. With the aim of fostering their awareness on private data leakage risk, some measures have been proposed that quantify the privacy risk of each user. However, these measures do not capture the objective risk of users since they assume that all user's direct social connections are close (thus trustworthy) friends. Since this assumption is too strong, in this paper we propose an alternative approach: each user decides which friends are allowed to see each profile item/post and our privacy score is defined accordingly. We show that it can be easily computed with minimal user intervention by leveraging an active learning approach. Finally, we validate our measure on a set of real Facebook users.