Research Reveals Privacy Risks in Interactive Targeted Ads

This research investigates the privacy implications of interactive targeted advertising systems, specifically focusing on how user attributes can be inferred from their interactions. The study models a scenario where an advertiser receives observations tied to individual users, rather than just aggregate reports, when an interaction is linked to the specific campaign that elicited it. This channel is conceptualized as a noisy oracle for inferring user attributes. The model meticulously separates various components: targeting predicates, ad exposure, user interaction, and the disclosure of information to the advertiser. This distinction helps to understand the gaps between a user's eligibility for an ad, its actual delivery, their interaction with it, and what information becomes visible to the advertiser. To evaluate these privacy risks, the researchers developed a reproducible benchmark using synthetic populations with known sensitive labels and generated campaign semantics. Through simulations, they compared the effectiveness of different attack methods, including Bayesian, supervised, positive and unlabeled, and adaptive attacks, under various campaign and disclosure definitions. The findings indicate that repeated campaigns with identity exposure can yield measurable, though bounded, inference signals, with Bayesian and supervised attacks achieving significant AUC scores. Crucially, the research identifies disclosure policy as the most potent control mechanism, suggesting that aggregate reporting, type filtering, and randomized disclosure can effectively reduce the signal available for attribute inference.