FairAudit-Platform: Building an Auditable AI Fairness Evaluation System

FairAudit-Platform is an open-source AI model fairness auditing platform developed based on Microsoft Fairlearn, Streamlit, and Scikit-learn. It supports evaluation of DPD (Demographic Parity Difference) and EOD (Equalized Odds Difference) metrics, bias mitigation workflows, and compliance with OWASP ML Top 10 and NIST AI RMF standards. The platform aims to help data scientists identify, quantify, and mitigate model biases, and is suitable for high-risk fields such as finance, healthcare, and recruitment.

With the widespread application of machine learning models in high-risk fields like finance, healthcare, and recruitment, the issue of algorithmic bias has become increasingly prominent—models may amplify social inequities due to historical biases in training data. Traditional evaluations only focus on average performance metrics such as accuracy, ignoring performance differences across different groups. Therefore, establishing a systematic AI fairness auditing mechanism has become a necessary condition for responsible AI development.

FairAudit-Platform is an open-source AI fairness auditing platform that provides an end-to-end workflow from model upload to bias mitigation. The core tech stack includes:

• Microsoft Fairlearn: Fairness evaluation and mitigation library
• Streamlit: Interactive web interface framework
• Scikit-learn: Foundation for machine learning training and evaluation

The platform supports two core fairness metrics:

1. Demographic Parity Difference (DPD)：Measures the difference in prediction distribution across different sensitive groups. The formula is DPD = P(Ŷ=1|A=0) - P(Ŷ=1|A=1), with an ideal value close to 0.
2. Equalized Odds Difference (EOD)：Measures the difference in True Positive Rate (TPR) and False Positive Rate (FPR) across different groups. The formula is EOD = max{|TPR(A=0)-TPR(A=1)|, |FPR(A=0)-FPR(A=1)|}, with an ideal value of 0.

The platform's features include:

1. Model Upload and Configuration: Supports uploading Scikit-learn models (e.g., pickle format) and specifying sensitive attributes and target variables.
2. Test Dataset Analysis: Automatically calculates basic statistical information for sensitive groups.
3. Fairness Visualization: Generates prediction distribution comparisons, confusion matrix heatmaps, metric radar charts, etc.
4. Bias Mitigation: Provides strategies such as reweighting, threshold optimization, and post-processing calibration.

Compliance: Adheres to OWASP ML Top 10 (e.g., detecting model theft, poisoning, drift) and NIST AI RMF (governance, mapping, measurement, management). Application Scenarios:

• Financial Credit: Detect group bias in loan approval processes
• Medical Diagnosis: Ensure consistent diagnostic accuracy across different groups
• Recruitment: Audit systematic bias in resume screening models

Technical Implementation: Modular architecture, including data layer (data processing with Pandas), model layer (Scikit-learn integration), fairness layer (Fairlearn encapsulation), visualization layer (Plotly + Streamlit), and workflow layer (state machine management). Limitations: Currently, it mainly supports binary classification tasks with tabular data, and does not cover multi-classification, regression, deep learning, unstructured data, etc.

Future Directions: Introduce causal fairness frameworks, support distributed evaluation for federated learning scenarios, and deeply integrate with MLOps platforms. Conclusion: FairAudit-Platform translates academic achievements into engineering practice, helping build fair models and support compliance audits. Fairness should be a necessary step for model deployment.