Xiya Shao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xiya Shao has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Ning Wang
Cyberspace Institute of Advanced Technology, Guangzhou University, Guangzhou, China
1 shared publication - Peng Tang
Shandong University, Qingdao, China
1 shared publication - Rui Chen
Harbin Engineering University, Harbin, China
1 shared publication - Shanqing Guo
Shandong University
1 shared publication
Latest Publications
- Open Access
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-26
Enhancing Local Differential Privacy Accuracy by Exploiting Inherent Uncertainty
Local differential privacy (LDP) is a standard for privacy preservation in the local setting: classical LDP mechanisms protect directly sensitive attributes by injecting uncertainty between users' true values and reported data. In many deployments, however, practitioners also choose to perturb attributes that are not inherently sensitive but are correlated with latent sensitive ones, using LDP as a system-level choice to limit attribute-inference risks. When such correlated attributes are naively treated as fully sensitive, the resulting mechanisms may satisfy a given privacy requirement but inject more noise than necessary, causing utility loss. This raises the challenge of quantifying how much uncertainty correlations already provide about sensitive attributes and using this to optimize perturbation under fixed privacy constraints. We address this challenge via inherent uncertainty, a metric that captures correlation-induced uncertainty between collected non-sensitive attributes and associated sensitive attributes. We also introduce a correlation-aware LDP framework that exploits this metric to recalibrate perturbation parameters, while providing the same ε-LDP guarantee for sensitive attributes as a treat-as-sensitive baseline. We further consider a hybrid scenario in which a single collected attribute contains both sensitive values and non-sensitive-but-correlated values, and propose a dual-phase perturbation method that provides differentiated protection and improves utility. Our mechanisms satisfy ε-LDP and the specified inference guarantees, and experiments on real datasets demonstrate their utility gains. Code is available at https://github.com/null944/enhanced-accuracy-in-ldp.
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-26
Open AccessEnhancing Local Differential Privacy Accuracy by Exploiting Inherent Uncertainty
Local differential privacy (LDP) is a standard for privacy preservation in the local setting: classical LDP mechanisms protect directly sensitive attributes by injecting uncertainty between users' true values and reported data. In many deployments, however, practitioners also choose to perturb attributes that are not inherently sensitive but are correlated with latent sensitive ones, using LDP as a system-level choice to limit attribute-inference risks. When such correlated attributes are naively treated as fully sensitive, the resulting mechanisms may satisfy a given privacy requirement but inject more noise than necessary, causing utility loss. This raises the challenge of quantifying how much uncertainty correlations already provide about sensitive attributes and using this to optimize perturbation under fixed privacy constraints. We address this challenge via inherent uncertainty, a metric that captures correlation-induced uncertainty between collected non-sensitive attributes and associated sensitive attributes. We also introduce a correlation-aware LDP framework that exploits this metric to recalibrate perturbation parameters, while providing the same ε-LDP guarantee for sensitive attributes as a treat-as-sensitive baseline. We further consider a hybrid scenario in which a single collected attribute contains both sensitive values and non-sensitive-but-correlated values, and propose a dual-phase perturbation method that provides differentiated protection and improves utility. Our mechanisms satisfy ε-LDP and the specified inference guarantees, and experiments on real datasets demonstrate their utility gains. Code is available at https://github.com/null944/enhanced-accuracy-in-ldp.