Balancing Privacy and Utility: Evaluating Distributional Shifts and Accuracy in Differentially Private Synthetic Breached Data

With increasing reliance on data-driven technologies, it is fundamental to ensure the privacy of individuals in datasets. This paper investigates the distributional shift introduced by differential privacy using a synthetically generated dataset simulating leaked personal information. We apply the Laplace mechanism to a hypothetical hotel booking adversarial disclosure scenario and analyze the impact of varying privacy budget (ε) and sensitivity (∆f ) parameters across 6,120 combinations. Through Jensen-Shannon Distance and Mean Absolute Percentage Error metrics, we quantify distributional shifts and accuracy degradation. Our findings reveal that attributes with higher entropy experience greater shift under noise addition, contributing with parameter tuning strategies for protecting sensitive data while preserving its analytical value.