Adaptive real-time key–value stream statistics and publishing with local differential privacy

Abstract In mobile crowdsensing (MCS) systems, real-time data monitoring and interpretation enable the provision of various services to users. As a pervasive NoSQL data model, key–value data appear in various MCS applications. However, indiscriminate collection and analysis of key–value data will lead to serious privacy risks for users. As for privacy-preserving stream publishing, existing differential privacy-based mechanisms are primarily oriented toward simple SQL data models, resulting in a lack of effective privacy protection solutions for key–value streams. To address this gap, this paper investigates novel and efficient privacy-preserving mechanisms for adaptive key–value stream publishing. Specifically, we employ standard local differential privacy techniques to perturb key–value streams before uploading to the server, thus providing strong local privacy guarantees for each user. We leverage adaptive population division methods to circumvent privacy budget division, greatly mitigating utility loss. In addition, we propose a dynamic user reweighting mechanism that updates user weights by adapting to key–value stream changes, thereby achieving a balanced trade-off between privacy and data utility. Experimental evaluations on multiple datasets indicate that the proposed mechanisms outperform state-of-the-art methods in data utility while providing strong privacy protection, demonstrating their potential value for privacy-preserving key–value stream processing in MCS applications.