A significant portion of the modern internet is funded by commercial return from customised content such as advertising where user interests are learned from users' online behaviour and used to display personalised content. Privacy becomes a concern when personalisation reveals evidence of learning about sensitive topics a user would rather keep private. Examples of potentially sensitive topics we consider include health, finance and sexual orientation. In this thesis we develop novel technologies allowing users to improve control over their personal privacy. We consider three aspects of privacy protection here: i) detecting evidence of unwanted profiling, ii) assessing the potential impact of a threat, and, iii) a flexible framework to help users to take control the flow of information used in personalisation. We model online systems as black-box adversaries with unknown internal workings but with an objective to maximise commercial utility. In a black-box environment absolute measures of privacy are problematic and so our formalism builds on a notion of privacy relative to a baseline. The relative models we develop have the advantage of being learn-able from observation of the black-box system and so can be readily implemented as practical technologies for privacy threat detection, analysis and privacy defence which we validate against data from well-known, real-world online systems.

APPROVED