Addressing Sensitivity Distinction in Local Differential Privacy: A General Utility-Optimized Framework

Local Differential Privacy (LDP) is widely employed to address privacy concerns in data collection. Nevertheless, the LDP model ignores the sensitivity distinction, as it regards all personal data equally sensitive, leading to excessive obfuscation and the loss of utility. Utility-optimized LDP (ULDP) aims to mitigate this issue. However, existing ULDP mechanisms address sensitivity distinction in only a limited subset of LDP mechanisms. To systematically address sensitivity distinction in the LDP model, we propose the General LDP-to-ULDP Transformation Framework. This framework can convert any LDP mechanism into its corresponding ULDP mechanism while preserving key properties such as order-optimality and unbiased estimation. Then, we present the pure ULDP framework, which generalizes a class of ULDP mechanisms with strong performance guarantees. We develop a universal aggregation and utility analysis method applicable to all pure ULDP mechanisms, facilitating the analysis, comparison, and optimization of different ULDP mechanisms. After that, we transform three widely-used LDP mechanisms into their ULDP counterparts (uSS, uUE and uLH). We theoretically demonstrate that our proposed mechanisms exceed existing ULDP mechanisms in data utility and communication costs. Specifically, our uSS, uUE and uLH match the minimax risk lower bound within the ULDP framework. We also identify the optimal mechanism for various usage scenarios. Finally, we conduct experiments on both real and synthetic datasets, showing that uUE and uLH achieve the lowest Mean Squared Error (MSE) when size of sensitive dataset is large, and uSS consistently achieves the lowest MSE.