visionai-fairness 0.1.0

AI Fairness Auditing Toolkit — detect, measure, and fix bias in ML models and datasets

VisionAI — AI Fairness Auditing Toolkit

Detect, measure, and fix bias in ML models and datasets.

VisionAI audits datasets and models for fairness violations including disparate impact, proxy discrimination, feature laundering, intersectional bias, and more. It maps findings to legal regulations (EU AI Act, EEOC, GDPR) and provides actionable fix recommendations.

Installation

pip install visionai

With SHAP explainability support:

pip install visionai[shap]

Quick Start

from visionai import FairnessAudit

audit = FairnessAudit(
    data="loan_data.csv",
    label_col="approved",
    positive_label="1",
    protected_cols=["gender", "race"],
    model="model.joblib",           # optional
    domain="Financial Lending",
)

results = audit.run()
print(results.fairness_score)       # 62
print(results.letter_grade)         # "C"
print(results.summary())
results.to_json("audit_report.json")

Features

Core Analysis

Module	Description
Data Bias Scanner	Disparate Impact ratio, Statistical Parity Difference, label skew
Model Bias Evaluator	Counterfactual perturbation testing, Equalized Odds (FPR/FNR)
Proxy Detector	Cramér's V + Eta-squared to find indirect discrimination channels
Feature Laundering	GradientBoosting reconstruction attack on protected attributes
Intersectional Audit	Pairwise protected attribute DI with significance thresholds
Explainability	SHAP values per demographic group, disparity detection
Severity Scorer	Weighted 0-100 score with letter grade (A-F)
Historical Harm	Estimate individuals harmed over deployment period
Flip Sensitivity	Decision boundary vulnerability analysis

Advanced Features (Phase 7)

Module	Description
Shadow Testing	Statistical synthetic profiles for missing demographic intersections
Adversarial Simulator	Minimum feature changes to flip a prediction
Red Team Mode	Worst-case bias search across all thresholds × demographic slices
Whistleblower Export	Anonymized reports with SHA-256 integrity hash
Model Comparison	Diff two audits — show improved/worsened metrics
Bias Origin Tracer	Data bias vs model bias — learned or amplified?

Compliance

Maps findings to EU AI Act, US EEOC, GDPR Article 22, India DPDP Act, UK Equality Act, and domain-specific regulations.

Granular Usage

Use individual modules directly:

from visionai import scan_data_bias, detect_proxies, detect_feature_laundering
import pandas as pd

df = pd.read_csv("data.csv")
bias = scan_data_bias(df, "hired", "1", ["gender", "race"])
proxies = detect_proxies(df, ["gender", "race"])
laundering = detect_feature_laundering(df, ["gender"], ["age", "income", "score"])

Shadow Testing

audit = FairnessAudit(data=df, label_col="approved", positive_label="1",
                       protected_cols=["gender", "race"], model=model)
shadow = audit.shadow_test()
print(shadow["missing_intersections"])
print(shadow["summary"]["flaggedCount"])

License

Apache 2.0 — See LICENSE