Towards optimal strategies for the management of online information and activity: privacy and utility tradeoffs

Abstract

The unprecedented growth of big-data applications suggests that there is a growing competition in the technological world to collect and harness tremendous amounts of user information. Tech companies and other online service providers are always seeking to enhance the quality of their products and services by collecting massive amounts of information from their user base. The collected data is typically used by the service providers to enhance the utility of the services. For instance, e-commerce services use the information about a user's purchases to recommend new products that may be of interest to the user. Similarly, streaming services use a user's ratings of various movies to recommend new and potentially interesting movies to the user. Unfortunately, the pursuit of utility often entails the loss of user privacy as the collected information often reveals sensitive information about the users, often through correlations not immediately apparent at the surface. This is aggravated by the fact that service providers often share, and even sell, their customers' information with third parties, which makes protecting the users' private information ever so critical. This dissertation seeks to address two important privacy problems. First, ensuring user privacy is not a trivial problem. At one end, service providers need customers' information to offer customized contents and personalized recommendations. The utility provided to a user is therefore positively correlated with the amount and the accuracy of the information that the user discloses to the service provider. On the other end, the collected information can be subject to inference attacks that reveal various private attributes of the user such as their income, race, political affiliation, and sexual orientation. The privacy of the user is therefore negatively correlated with the amount of disclosed information. The problem, as such, naturally manifests as a privacy-utility tradeoff problem. In this dissertation, we develop models to capture the precise notions of privacy and utility and design privacy mechanisms that maximize the utility of the disclosed information while limiting the privacy leakage. The second problem that this dissertation seeks to address is of extreme relevance: given users' tendency to continuously disclose their personal information, as in the case of social media, modeling the privacy leakage over time is paramount to devising privacy mechanisms that limit the accumulated leakage. Further, there is a natural concern regarding the effect of our current online activities on our future privacy. Modeling the problem is extremely intricate as capturing future privacy is not trivial given the inherent uncertainties surrounding the future. In this dissertation, we capture the dynamics of privacy leakage over time using a probabilistic framework. Via experimental evaluations, we demonstrate that there exist multiple promising strategies that a user can utilize to limit their future privacy leakage while maximizing their perceived utility over time.