String Discovery for Private Analytics

A number of research systems enable analysts to aggregate user data that is distributed across user devices while preventing online tracking and providing users with differential privacy guarantees. These systems rely on pre-defined string values to release relevant user data in a controlled fashion. Unfortunately, many string values (e.g., tags in a photo application) may not be easily predicted. Existing private aggregation systems that can be used to discover strings for private analytics purposes exhibit serious shortcomings, such as heavy client-side operations and an inability to deal with malicious clients supplying incorrect data. In this paper, we present a practical and privacy-preserving string discovery system that provides analysts with previously unknown strings and limits the effects of malicious clients, while supporting a variety of user devices with varying computation and bandwidth resources. To achieve these goals, our system employs the exclusive-OR (XOR) operation as its crypto primitive, and utilizes a novel method to determine the equivalence of two XOR-encrypted strings without revealing them. We present our design, analyze its privacy properties and evaluate its feasibility. Our results show that our system outperforms the closest system by several orders of magnitude for client-side computations and one order of magnitude for server-side computations.