Preserving privacy in geo-targeted advertising

Targeted advertising has inherent privacy risks: ad providers aim to maximize the information inferred about the users in order to increase their click-through ratio. This in turn leads to long-term privacy risks for the users as their information is traded among ad agencies and unknown third parties. In this paper we focus on the privacy associated with information such as location and income, and their relationships with the ads served to users. We first show
the possibility for an attacker to use topic modelling and machine learning techniques on ads served to a user as a means to accurately infer their location and income band within a city. We then attempt to reduce this risk of inference using obfuscation, or hiding in plain sight. The idea is to hide the targeted ads, not by relying on encryption, rather by producing just enough noise such that an attacker cannot distinguish between the actual ads served and the ads which
are just noise. Our results are promising and demonstrate that efforts such as TrackMeNot can help advertising and users achieve a balance between targeted advertising and location privacy exposure.