# FairProof : Confidential and Certifiable Fairness for Neural Networks

**Authors:** Chhavi Yadav, Amrita Roy Chowdhury, Dan Boneh, Kamalika Chaudhuri  
**arXiv:** [2402.12572](https://arxiv.org/abs/2402.12572)  
**Date:** 2024-02-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.12572v1/

## Key points

- Machine learning models are increasingly used in societal applications
- Concerns about legal and privacy issues necessitate model confidentiality
- Fairness skepticism among consumers impacted by model predictions
- FairProof system introduced to address fairness concerns using Zero-Knowledge Proofs (ZKPs)
- FairProof includes a fairness certification algorithm for fully-connected neural networks compatible with ZKPs and implemented using Gnark
- FairProof issues fairness certificates by inspecting model weights post-training for neural networks without specific training pipelines
- Classification of certification mechanisms for fairness into three categories: optimization programs, robustness-fairness connection leveraging mechanisms, and black-box query access learning theoretic approaches
- FairProof focuses on local Individual Fairness (IF) for neural networks using an iterative algorithm, demonstrating practical feasibility for small neural networks and tabular data
- Protocol enables model owners to issue publicly verifiable certificates while ensuring uniformity and confidentiality
- Experiments show practical feasibility of FairProof in enhancing transparency and trustworthiness in machine learning models used in societal applications

## Summary

Machine learning models are increasingly being utilized in various societal applications. However, concerns regarding legal and privacy issues often require that these models remain confidential. This has led to a growing skepticism about the fairness of these models among consumers who are impacted by their predictions. To address this issue, a system called FairProof has been proposed. <br>
<br>
FairProof utilizes Zero-Knowledge Proofs (ZKPs), a cryptographic primitive, to publicly verify the fairness of a model while still maintaining confidentiality. Additionally, a fairness certification algorithm tailored for fully-connected neural networks is introduced. This algorithm is compatible with ZKPs and implemented in the FairProof system using Gnark.<br>
<br>
In related work, previous research has focused on verifying fairness while preserving model confidentiality through third-party auditors or using cryptographic primitives to verify other properties of machine learning models. However, FairProof stands out by issuing fairness certificates by inspecting model weights post-training for neural networks without being tied to specific training pipelines.<br>
<br>
Furthermore, there is a classification of certification mechanisms for fairness into three categories: optimization programs, robustness-fairness connection leveraging mechanisms like adversarial training, and black-box query access learning theoretic approaches. FairProof specifically focuses on local Individual Fairness (IF) for neural networks using an iterative algorithm and demonstrates practical feasibility for small neural networks and tabular data.<br>
<br>
In conclusion, FairProof offers a protocol that enables model owners to issue publicly verifiable certificates while ensuring model uniformity and confidentiality. The experiments conducted showcase the practical feasibility of FairProof in enhancing transparency and trustworthiness in machine learning models used in societal applications. The use of ZKPs as a general-purpose tool holds promise beyond fairness considerations in machine learning systems.