Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quantitative Auditing of AI Fairness with Differentially Private Synthetic Data

Chih-Cheng Rex Yuan, Bow-Yaw Wang

TL;DR

The paper tackles privacy risks in fairness auditing of AI systems by introducing a framework that uses differentially private synthetic data to audit fairness. It builds on a margin-based, Markov random field approach to generate synthetic datasets with Rényi differential privacy guarantees and evaluates fairness metrics using an open-source checker. Empirical results on the Adult, COMPAS, and Diabetes datasets show that synthetic data can preserve the general fairness properties of real data, with average metric differences typically below $0.1$, though some sufficiency-related measures are harder to approximate. The work demonstrates a practical, privacy-preserving path for auditing AI fairness in sensitive domains, while also outlining limitations and directions for scalability and policy integration.

Abstract

Fairness auditing of AI systems can identify and quantify biases. However, traditional auditing using real-world data raises security and privacy concerns. It exposes auditors to security risks as they become custodians of sensitive information and targets for cyberattacks. Privacy risks arise even without direct breaches, as data analyses can inadvertently expose confidential information. To address these, we propose a framework that leverages differentially private synthetic data to audit the fairness of AI systems. By applying privacy-preserving mechanisms, it generates synthetic data that mirrors the statistical properties of the original dataset while ensuring privacy. This method balances the goal of rigorous fairness auditing and the need for strong privacy protections. Through experiments on real datasets like Adult, COMPAS, and Diabetes, we compare fairness metrics of synthetic and real data. By analyzing the alignment and discrepancies between these metrics, we assess the capacity of synthetic data to preserve the fairness properties of real data. Our results demonstrate the framework's ability to enable meaningful fairness evaluations while safeguarding sensitive information, proving its applicability across critical and sensitive domains.

Quantitative Auditing of AI Fairness with Differentially Private Synthetic Data

TL;DR

The paper tackles privacy risks in fairness auditing of AI systems by introducing a framework that uses differentially private synthetic data to audit fairness. It builds on a margin-based, Markov random field approach to generate synthetic datasets with Rényi differential privacy guarantees and evaluates fairness metrics using an open-source checker. Empirical results on the Adult, COMPAS, and Diabetes datasets show that synthetic data can preserve the general fairness properties of real data, with average metric differences typically below , though some sufficiency-related measures are harder to approximate. The work demonstrates a practical, privacy-preserving path for auditing AI fairness in sensitive domains, while also outlining limitations and directions for scalability and policy integration.

Abstract

Fairness auditing of AI systems can identify and quantify biases. However, traditional auditing using real-world data raises security and privacy concerns. It exposes auditors to security risks as they become custodians of sensitive information and targets for cyberattacks. Privacy risks arise even without direct breaches, as data analyses can inadvertently expose confidential information. To address these, we propose a framework that leverages differentially private synthetic data to audit the fairness of AI systems. By applying privacy-preserving mechanisms, it generates synthetic data that mirrors the statistical properties of the original dataset while ensuring privacy. This method balances the goal of rigorous fairness auditing and the need for strong privacy protections. Through experiments on real datasets like Adult, COMPAS, and Diabetes, we compare fairness metrics of synthetic and real data. By analyzing the alignment and discrepancies between these metrics, we assess the capacity of synthetic data to preserve the fairness properties of real data. Our results demonstrate the framework's ability to enable meaningful fairness evaluations while safeguarding sensitive information, proving its applicability across critical and sensitive domains.
Paper Structure (18 sections, 1 theorem, 11 equations, 1 figure, 9 tables)

This paper contains 18 sections, 1 theorem, 11 equations, 1 figure, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Fairness Measures
  5. Rényi differential privacy
  6. Differentially Private Synthetic Data
  7. Motivation
  8. Methodology
  9. Data Synthesis
  10. Fairness Checking
  11. Experiments
  12. Adult Income Dataset
  13. COMPAS Dataset
  14. Diabetes Dataset
  15. Evaluation
  16. ...and 3 more sections

Key Result

theorem 1

The Gaussian Mechanism satisfies $(\alpha, \alpha \frac{\Delta^2_f}{2 \sigma^2})$-RDP, where $\Delta_f$ denotes the sensitivitydwork2014algorithmic of $f$, which is defined as the maximum $L^2$-norm difference in the output of $f$

Figures (1)

  • Figure 1: Roles of each party in the COMPAS example.

Theorems & Definitions (3)

  • definition 1: Gaussian Mechanismdwork2014algorithmic
  • definition 2: Rényi Differential Privacy (RDP)
  • theorem 1: RDP of the Gaussian Mechanismfeldman2018privacymironov2017renyi