Abstract:As computation over sensitive data has been an important goal in recent years, privacy-preserving data analysis has gradually attracted more and more attention. Among various mechanisms, differential privacy has been widely studied due to its formal privacy guarantees for data analysis. As one of the most important issues, the crucial trade-off between the strength of privacy guarantee and the effect of analysis accuracy is highly concerned among the researchers. Existing theories for this issue consider that the analyst should first choose a privacy requirement and then attempt to maximize the utility. However, as differential privacy is gradually deployed in practice, a gap between theory and practice comes out: in practice, product requirements often impose hard accuracy constraints, and privacy (while desirable) may not be the over-riding concern. Thus, it is usually that the requirement of privacy guarantee is adjusted according to the utility expectation, not the other way around. This gap raises the question of how to provide maximum privacy guarantee for data analysis due to a given accuracy requirement. In this paper, we focus our attention on private Empirical Risk Minimization (ERM), which is one of the most commonly used data analysis method. We take the first step towards solving the above problem by theoretically exploring the effect of ϵ (the parameter of differential privacy that determines the strength of privacy guarantee) on utility of the learning model. We trace the change of utility with modification of ϵ and reveal an established relationship between ϵ and utility. We then formalize this relationship and propose a practical approach for estimating the utility under an arbitrary value of ϵ. Both theoretical analysis and experimental results demonstrate high estimation accuracy and broad applicability of our approach in practical applications. As providing algorithms with strong utility guarantees that also give privacy when possible becomes more and more accepted, our approach would have high practical value and may be likely to be adopted by companies and organizations that would like to preserve privacy but are unwilling to compromise on utility.

Just Keep Your Concerns Private: Guaranteeing Heterogeneous Privacy and Achieving High Availability for ERM Algorithms

Towards Practical Differential Privacy in Data Analysis: Understanding the Effect of Epsilon on Utility in Private ERM

DPMLBench: Holistic Evaluation of Differentially Private Machine Learning

A Differentially Private Framework for the Dynamic Heterogeneous Redundant Architecture System in Cyberspace

Accuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained ERM

Not one but many Tradeoffs: Privacy Vs. Utility in Differentially Private Machine Learning

Data Heterogeneity Differential Privacy: From Theory to Algorithm.

Differential Privacy: An Economic Method for Choosing Epsilon

Heterogeneous Differential-Private Federated Learning: Trading Privacy for Utility Truthfully.

Mitigating Disparate Impact on Model Accuracy in Differentially Private Learning.

Meeting Utility Constraints in Differential Privacy: A Privacy-Boosting Approach

Optimizing Batch Linear Queries under Exact and Approximate Differential Privacy

Differentially Private ERM Based on Data Perturbation

Noise-Aware Algorithm for Heterogeneous Differentially Private Federated Learning

When Homomorphic Cryptosystem Meets Differential Privacy: Training Machine Learning Classifier with Privacy Protection.

Differentially private regression analysis with dynamic privacy allocation

Evaluating Differentially Private Machine Learning in Practice

Generalized PTR: User-Friendly Recipes for Data-Adaptive Algorithms with Differential Privacy

Rethinking Improved Privacy-Utility Trade-off with Pre-existing Knowledge for DP Training

Privacy Budgeting for Growing Machine Learning Datasets

On the Importance of Architecture and Feature Selection in Differentially Private Machine Learning