Abstract :
[en] Online social networks (OSNs) have been the most successful online applications during the past decade. Leading players in the business, including Facebook, Twitter and Instagram, attract a huge number of users. Nowadays, OSNs have become a primary way for people to connect, communicate and share life moments. Although OSNs have brought a lot of convenience to our life, users' privacy, on the other hand, has become a major concern due to the large amount of personal data shared online. In this thesis, we study users' privacy in social networks from two aspects, namely access control and information inference.
Access control is a mechanism, provided by OSNs, for users themselves to regulate who can view their resources. Access control schemes in OSNs are relationship-based, i.e., a user can define access control policies to allow others who are in a certain relationship with him to access his resources. Current OSNs have deployed multiple access control schemes, however most of these schemes do not satisfy users' expectations, due to expressiveness and usability.
There are mainly two types of information that users share in OSNs, namely their activities and social relations. The information has provided an unprecedented chance for academia to understand human society and for industry to build appealing applications, such as personalized recommendation. However, the large quantity of data can also be used to infer a user's personal information, even though not shared by the user in OSNs.
This thesis concentrates on users' privacy in online social networks from two aspects, i.e., access control and information inference, it is organized into two parts.
The first part of this thesis addresses access control in social networks from three perspectives. First, we propose a formal framework based on a hybrid logic to model users' access control policies. This framework incorporates the notion of public information and provides users with a fine-grained way to control who can view their resources. Second, we design cryptographic protocols to enforce access control policies in OSNs. Under these protocols, a user can allow others to view his resources without leaking private information. Third, major OSN companies have deployed blacklist for users to enforce extra access control besides the normal access control policies. We formally model blacklist with the help of a hybrid logic and propose efficient algorithms to implement it in OSNs.
The second part of this thesis concentrates on the inference of users' information in OSNs, using machine learning techniques. The targets of our inference are users' activities, represented by mobility, and social relations. First, we propose a method which uses a user's social relations to predict his locations. This method adopts a user's social community information to construct the location predictor, and perform the inference with machine learning techniques. Second, we focus on inferring the friendship between two users based on the common locations they have been to. We propose a notion namely location sociality that characterizes to which extent a location is suitable for conducting social activities, and use this notion for friendship prediction. Experiments on real life social network datasets have demonstrated the effectiveness of our two inferences.
