Rank-Based Inference over Web Databases

A fundamental virtue of social media is to build virtual communities between users. As such, it is no surprise that almost all social media sites provide web interfaces for the search and/or recommendation of other users who share similar attributes, interests, etc. with results being the top-k users selected according to a ranking function. Our studies of real-world websites unveil a novel yet serious privacy leakage caused by the design of such interfaces and ranking functions. Specifically, we find that many such websites feature private attributes only visible to a user him/herself, but not to other users (and therefore will not be visible in the query answer). Nonetheless, some websites also take into account such private attributes in the design of the ranking function, understandably for improving the effectiveness of search/recommendation. While the conventional belief might be that tuple ranks alone are not enough to reveal the private attribute values, our investigation shows that this is not the case in reality. Specifically, we define a novel problem of rank based inference, and introduce a taxonomy of the problem space according to two dimensions, (1) the type of query interfaces widely used in practice and (2) the capability of adversaries. For each subspace of the problem, we develop a novel technique which either guarantees the successful inference of private attributes, or (when such an inference is provably infeasible in the worst-case scenario) accomplishes such an inference attack for a significant portion of real-world tuples. We demonstrate the effectiveness and efficiency of our techniques through theoretical analysis and extensive experiments over realworld datasets, including successful online attacks over popular social media such as Amazon Goodreads and Catch22dating.