🎯 What is privacylens?

Most ML engineers don't know if their model is leaking private training data. privacylens audits it.

from privacylens import audit

report = audit(model, X_train, y_train, X_test)
report.summary()

┌─────────────────────────────────────────────────────────────┐
│             🔍 privacylens — Privacy Audit Report            │
├──────────────────────────────┬────────────┬─────────────────┤
│ Check                        │ Score      │ Risk            │
├──────────────────────────────┼────────────┼─────────────────┤
│ Membership Inference Attack  │ 0.087      │ LOW             │
│ PII Leakage Detection        │ 0.000      │ LOW             │
│ Model Inversion Risk         │ 0.042      │ LOW             │
└──────────────────────────────┴────────────┴─────────────────┘

Model: RandomForestClassifier
Overall Risk: LOW

• MIA advantage score: 0.087 — model shows low Membership Inference vulnerability.
• PII leakage score: 0.000 — model shows low PII Leakage vulnerability.
• Model Inversion score: 0.042 — model shows low Model Inversion vulnerability.

✨ Features

• 🕵️ Membership Inference Attack (MIA) — Detect if an attacker can identify training records using shadow model estimation (Shokri et al., 2017)
• 🔎 PII Leakage Detection — Detect sensitive PII (Emails, SSNs, Credit Cards, Phones, IPs) memorized in predictions or samples
• 🔄 Model Inversion Risk Scorer — Evaluate feature reconstructability risk from output confidence probabilities (Fredrikson et al., 2015)
• 🌐 Framework agnostic — Works with scikit-learn, XGBoost, and PyTorch models
• 🎨 Beautiful terminal output — Rich colour-coded risk tables with LOW / MEDIUM / HIGH classification
• 🤖 CLI + Python API — Use in scripts or integrate into CI/CD pipelines (privacylens audit)
• 📊 JSON output — Machine-readable results for dashboards and reporting (--output json)

Coming in future releases:

• 📄 HTML compliance report export (Jinja2) for GDPR/HIPAA auditing
• 🤗 HuggingFace Transformers LLM adapter

📦 Installation

# Base install (scikit-learn models)
pip install privacyaudit

# With PyTorch support
pip install "privacyaudit[torch]"

# With XGBoost support
pip install "privacyaudit[xgboost]"

# Everything
pip install "privacyaudit[all]"

Note: The PyPI package is privacyaudit. Import in Python as from privacylens import audit.

🚀 Quick Start

from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from privacylens import audit

# Train a model
X, y = make_classification(n_samples=1000, n_features=20, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

model = RandomForestClassifier(n_estimators=100)
model.fit(X_train, y_train)

# Audit it for privacy vulnerabilities (MIA + PII Leakage + Model Inversion)
report = audit(model, X_train, y_train, X_test, y_test)
report.summary()

# Get audit results as dict (for JSON logging or API responses)
print(report.to_dict())

🖥️ CLI Usage

# Audit a saved model file
privacylens audit model.pkl train.csv test.csv

# Output JSON for CI/CD integration
privacylens audit model.pkl train.csv test.csv --output json

# Skip MIA check in fast pipelines
privacylens audit model.pkl train.csv test.csv --no-mia

🏗️ Architecture

privacylens/
├── src/privacylens/
│   ├── __init__.py         # Public API: audit(), AuditReport, Auditors
│   ├── auditor.py          # Core orchestrator
│   ├── attacks/
│   │   ├── membership.py   # MIA engine (shadow model + attack classifier)
│   │   └── inversion.py    # Model Inversion Risk Auditor (Fredrikson et al.)
│   ├── leakage/
│   │   └── pii.py          # PII Leakage Auditor (Regex + Severity Weighting)
│   └── cli.py              # Click CLI
└── tests/
    ├── test_auditor.py
    ├── test_membership.py
    ├── test_pii_leakage.py
    └── test_inversion.py

📖 Risk Score Interpretation

🤝 Contributing

See CONTRIBUTING.md. All contributions welcome!

📄 License

MIT — see LICENSE.