[Paper Review] PrivFair: a Library for Privacy-Preserving Fairness Auditing

Sikha Pentyala, David Melanson+2 more|arXiv (Cornell University)|Feb 8, 2022

Privacy-Preserving Technologies in Data4 citations

TL;DR

PrivFair is a novel library that enables privacy-preserving fairness auditing of machine learning models using Secure Multiparty Computation (MPC), allowing model owners and auditors to evaluate fairness metrics—such as demographic parity and equal opportunity—without disclosing sensitive model parameters or audit data. The system supports both 2PC and 3PC threat models, ensuring confidentiality while maintaining accuracy and scalability across tabular and image data.

ABSTRACT

Machine learning (ML) has become prominent in applications that directly affect people's quality of life, including in healthcare, justice, and finance. ML models have been found to exhibit discrimination based on sensitive attributes such as gender, race, or disability. Assessing if an ML model is free of bias remains challenging to date, and by definition has to be done with sensitive user characteristics that are subject of anti-discrimination and data protection law. Existing libraries for fairness auditing of ML models offer no mechanism to protect the privacy of the audit data. We present PrivFair, a library for privacy-preserving fairness audits of ML models. Through the use of Secure Multiparty Computation (MPC), PrivFair protects the confidentiality of the model under audit and the sensitive data used for the audit, hence it supports scenarios in which a proprietary classifier owned by a company is audited using sensitive audit data from an external investigator. We demonstrate the use of PrivFair for group fairness auditing with tabular data or image data, without requiring the investigator to disclose their data to anyone in an unencrypted manner, or the model owner to reveal their model parameters to anyone in plaintext.

Motivation & Objective

• To address the critical gap in fairness auditing tools that simultaneously protect model privacy and sensitive audit data.
• To enable fair evaluation of proprietary ML models without requiring disclosure of model weights or audit data to external parties.
• To support real-world fairness auditing scenarios in high-stakes domains like healthcare, finance, and criminal justice where data and model confidentiality are legally and ethically essential.
• To provide a practical, extensible framework that integrates with existing fairness metrics and supports diverse model architectures.

Proposed method

• PrivFair uses Secure Multiparty Computation (MPC) to enable joint computation over encrypted shares of model parameters and audit data, ensuring neither party reveals their inputs in plaintext.
• The system operates in a 2PC model where the model owner and auditor compute fairness metrics directly, or in a 3PC model where untrusted cloud servers perform the computation on encrypted shares.
• All computations are performed over a finite integer ring 𝕫_q to ensure compatibility with MPC protocols and numerical stability.
• The library implements MPC protocols for core fairness metrics including demographic parity, equal opportunity, equalized odds, and subgroup accuracy.
• It supports both tabular and image data by integrating with existing MPC-based inference protocols for models like logistic regression and ConvNets.
• The architecture allows asynchronous execution in 3PC mode, enabling auditors to query pre-shared encrypted models from cloud servers without ongoing interaction.

Experimental results

Research questions

• RQ1 Can fairness auditing of proprietary ML models be performed without disclosing sensitive model parameters or audit data?
• RQ2 How can Secure Multiparty Computation be effectively applied to privacy-preserving fairness evaluation in real-world ML deployment scenarios?
• RQ3 What is the performance overhead of using MPC for fairness auditing, and is it acceptable in practical, privacy-sensitive applications?
• RQ4 Can the same MPC infrastructure support both model inference and fairness metric computation in a unified, extensible framework?
• RQ5 How does the system scale across different model architectures and data types, including image and tabular data?

Key findings

• PrivFair successfully enables privacy-preserving fairness auditing of both tabular and image data using MPC, with no disclosure of model parameters or audit data in plaintext.
• The system achieves accurate fairness metric computation—e.g., demographic parity of 0.24 for males and 0.19 for females—while maintaining end-to-end confidentiality.
• Runtime for fairness auditing is dominated by MPC-based inference, with total execution time of 229.27 seconds for image classification, primarily due to secure inference costs.
• The performance overhead is justified in fairness auditing contexts, where longer runtimes are acceptable compared to real-time inference applications.
• The library scales linearly with data size and model complexity, consistent with MPC protocol behavior, and supports both 2PC and 3PC deployment models.
• PrivFair is extensible to additional fairness metrics such as treatment equality and predictive equality, and supports diverse model types including decision trees, random forests, and neural networks.