OKV: Optimized Key-Value Data Collection with Local Differential Privacy

Abstract

Local differential privacy (LDP), where each user obfuscates their data locally before sending it to an untrustworthy data collector, provides a strict privacy guarantee for users' sensitive data. However, the existing key-value data collection mechanisms based on the LDP assume that all keys are equally sensitive, which leads to excessive protection and thus loss of utility. To address the reduced utility caused by overprotection, we introduce the notion of key-value data utility-optimized LDP (KV-ULDP), which only offers a basic LDP-equivalent privacy guarantee for sensitive keys and all values. Subsequently, we design a new framework, named optimized key-value data collection (OKV) with LDP, which satisfies the KV-ULDP with high utility while keeping secret for each user. We instantiate the OKV framework by using OKV-UE (based on Unary Encoding) and OKV-GRR (based on Generalized Randomized Response) mechanisms. The OKV-UE is effective with a large number of key types, and OKV-GRR works well under high privacy budgets. The theoretical analysis and the experiments on two real datasets show that our mechanisms outperform the existing key-value mechanisms with LDP in terms of utility.