Priv-S: Privacy-Sensitive Data Identification in Online Social Networks.

Privacy inference imposes a serious threat to user privacy in Online Social Networks (OSNs) as the vast amount of personal data and relationships in OSNs can be used not only to infer user privacy but also to enrich the training set of inference methods. Previous studies have mostly focused on privacy inference from the perspective of the adversary with the objective of tracking the accuracy of the inference results. However, countering privacy inference requires not only the analysis of the inference method but also the roles that data would play in the inference process, i.e., to determine the sensitivity of the data. To address this issue, in this paper, we propose a model for the identification of privacy-sensitive data in which we formulate the identification as an influence maximization problem to identify both the privacy-sensitive users and the privacy-sensitive attributes for privacy inference. In our model, a privacy-affected tree is first constructed based on the influence between users with respect to the concerned privacy. Then, privacy-sensitive users in the privacy-affected tree are identified along with their privacy-sensitive attributes based on their contribution to the inference result of the concerned privacy. Experiments show that the results of our proposed model can significantly affect the accuracy of privacy inference, which demonstrates that our model can identify the privacy-sensitive data. Meanwhile, the impact of privacy-sensitive users and privacy-sensitive attributes is analyzed to guide the design of effective privacy-enhancing technologies.