Differentially Private Mechanisms Using Enhanced Direction-Oriented Smooth Sensitivity

As the volume of social data increases, the development of privacy-preserving methods for information processing is becoming critically important. Numeric queries are fundamental tasks in this domain, and handling them while satisfying differential privacy is crucial for ensuring trustworthy data management. However, foundational differentially private mechanisms and concepts still have room for refinement. This study focuses on advancing the concept of smooth sensitivity, which is helpful for adding data-dependent noise. Recently, a new concept called direction-oriented smooth sensitivity (DOSS) was proposed to consider the noise direction as well; however, this concept has several unresolved issues and offers potential for further noise reduction. Therefore, we first improve the concept of direction-oriented local sensitivity (DOLS) to reduce the lower bound of the noise scales. We then redefine DOSS based on this improved DOLS. Concurrently, we revisit the conditions for satisfying E-differential privacy and use the resulting enhanced DOSS to construct a more sophisticated mechanism. Furthermore, we propose an unbiased mechanism in which the expected value of the output equals the input value. Experimental results focused on the publication of genome statistics demonstrate that the proposed mechanisms can achieve higher utility than existing mechanisms. Overall, this study represents a significant step toward realizing mechanisms that can add more realistic and reliable noise. The supplemental materials are available at httJ)S://2ithub.com/av0408IEnhancedDOSS.