| Social networks bring both opportunities and threats to the users. On one hand, social networks provide a platform for users to build online profiles, make connections with others beyond geographical boundaries, enjoy the "openness" of social networks to meet their intrinsic need of "self-presentation", explore and strengthen their interests by joining interested virtual communities, etc. Besides, users also can benefit from recommendations provided by recommender systems that are employed by the social network service providers. On the other hand, users' privacy faces significant threats from potential attacks on social networks. Privacy risks on social networks are mainly caused by the variety of information being voluntarily or accidentally shared by the users, as well as implicit information that could be inferred based on the social relations and affiliations. These two aspects of social networks have raised attention from both academia and industry. In the past few years, a lot of work has been devoted to designing effective and efficient social recommendations to provide more opportunities for the users, identifying potential privacy attack models, proposing privacy protection strategies on social networks. However, most of the work either studies the Online Social Networks (OSNs) that model the online social world of the users, or investigates Location Based Social Networks (LBSNs) that target on users' online activities. Neither of them captures the full scenario of users' social relations and activities in the real life.;In this dissertation, we study the opportunities and threats of social networks by linking both online and offline social worlds. Specifically, we target a special type of social networks, namely Event Based Social Networks (EBSNs) that couple both online and offline social relations, as well as both online and offline activities. Users in EBSNs formalize two types of social interactions, i.e., "virtual" online interactions when they communicate online, and "physical" offline interactions when they participate in some offline social events and have face-to-face communications. We conducted a thorough study regarding EBSNs. To provide effective online group recommendations and offline event recommendations on EBSNs, we proposed an iterative and interactive recommendation framework by interactively employing both online and offline social relations. We also identified some potential privacy attack/inference models in EBSNs, extensively simulated and analyzed their destructive effect on online group membership inference and offline event attendance inference. Finally, we designed a privacy-preserving social recommendation framework on EBSNs that can produce accurate online and offline social recommendations, at the same time provide differential privacy guarantees to the users on EBSNs. The effectiveness of the proposed frameworks and models has been verified on a large dataset collected from a typical EBSN in the real world, i.e., Meetup. |
