Privacy in social networks

Abstract

Participation in social networks has surpassed all other online activities to be one of the top internet activities. Information disclosure on social networks comes with the risk of this information falling into wrong hands leading to privacy problems. In this thesis, the problem of privacy in social networks is researched in two aspects. This thesis firstly finds solutions and proposes enhancements to the existing and known privacy leaks. Secondly, advanced ways of inferring user and friends' private information, utilizing user and his friends public information, are proposed.

With the current privacy settings not mapping to the user's mental model of privacy settings, the proposed privacy aware intended audience selection model assists user in finding the right audience for the information to be posted on social network. According to a recent finding more than half of the online adults use multiple social networking sites. User is assisted in reducing repetitive task of setting privacy across multiple social networks using proposed privacy scoring model in which scores once calculated can be used on all social networks that the user and his friends share. This thesis is the first to provide a detailed account of attribute inference abilities from the follower and following subnetworks of the user, which are partial views of the user's whole network. This work highlights the increased risk of user attribute inference, as a partial view of the user's network and a small percentage (20\%) of friend's public attributes can lead to inference of user's private attributes. In the research on health information disclosure, the research work demonstrates the need to understand the context of information before drawing general health disclosure trends, with the work also demonstrating the privacy leaks prevalent due to the disclosures from the user himself and from his friends on social networks. Utilizing only a single user's public information, containing his friends' comments, likes and wall posts, friends' response prediction model has been successful in predicting the response by friends on user's posts along with predicting friends' latent affinity towards topics of interest. This amounts to privacy leaks of friends. This thesis uses datasets of popular social networks such as Facebook, Google+ and Twitter in demonstrating the applicability of all the proposed models through extensive experiments.