PriView: Personalized Media Consumption Meets Privacy against Inference Attacks

PriView: Personalized Media Consumption Meets Privacy against Inference Attacks PriView is an interactive personalized video consumption system that protects user privacy while recommending relevant content. It provides transparency of privacy risk, control of privacy risk, and personalized recommendations. It implements an information-theoretic framework to enable a utility-aware privacy mapping that distorts a user’s video ratings to prevent attackers from inferring users’ personal attributes (such as age, gender, or political views), while maintaining the distorted ratings’ usefulness for recommendations. PriView uses convex optimization to create a probability mapping from actual ratings to distorted ratings that minimizes the distortion, subject to a privacy constraint. One practical challenge is scalability, when data comes from a large alphabet. Quantization combined with low-rank approximation of the rating matrix helps reduce the number of optimization variables. Evaluations showed that PriView can achieve perfect privacy with little change in recommendation quality. This article is part of a special issue on Security and Privacy on the Web.