Solving the Performance Issues of Epsilon Estimation Method in Differentially Private ERM: Analysis, Solution and Evaluation

The study of differential privacy has recently garnered significant attention due to its ability to provide rigorous privacy protection for data analysis tasks. However, researchers are grappling with the critical balance between ensuring strong privacy protection and maintaining the utility of the model. While existing theories recommend selecting a privacy requirement and optimizing utility accordingly, practical product requirements often have strict accuracy constraints. As a result, privacy considerations may not always take precedence. Consequently, privacy guarantees are often adjusted to align with the expected level of utility. This mismatch highlights the challenge of training a highly private model that also meets the desired level of utility. Li et al. recently introduced a practical method for estimating utility at any privacy budget level, which can aid in achieving the aforementioned objective. However, upon analysis, our team has uncovered performance deficiencies in this approach. In this paper, we comprehensively examine the performance issues of Li’s method, identify their underlying cause, and present effective solutions. Our experimental results demonstrate that our proposed solutions can greatly improve the performance of Li’s method in a highly efficient manner.