Privacy Synthesis Tutorial

Learn how to generate synthetic data with differential privacy guarantees. This tutorial covers HIPAA, GDPR, and CCPA compliance requirements.

Tutorial Goals

By the end of this tutorial, you will:

• Understand differential privacy concepts
• Generate data with mathematical privacy guarantees
• Validate privacy configurations before training
• Review privacy reports and compliance documentation
• Choose appropriate privacy parameters for your use case

Time Required: 25-35 minutes Difficulty: Intermediate Prerequisites: Basic synthesis tutorial completed

Differential Privacy Fundamentals

What is Differential Privacy?

Differential Privacy (DP) provides mathematical guarantees that individual records cannot be distinguished in analysis results. It adds carefully calibrated noise to ensure privacy while preserving statistical utility.

Key Concept: A DP algorithm guarantees that whether or not any individual's data is included, the output distribution remains statistically indistinguishable.

Privacy Parameters

Epsilon (ε) - Privacy Budget

• Lower values = Stronger privacy (more noise)
• Higher values = Weaker privacy (less noise, better utility)
• Typical range: 0.1 to 20.0

Delta (δ) - Failure Probability

• Probability that privacy guarantee fails
• Typically: 1/dataset_size or 1e-5 for large datasets
• Auto-calculated in Synthetic Data Studio

Privacy Levels by Use Case

Use Case	Epsilon Range	Privacy Level	Compliance
Clinical trials	0.1 - 1.0	Very Strong	HIPAA PHI
Healthcare data	1.0 - 5.0	Strong	GDPR Article 9
Customer data	5.0 - 10.0	Moderate	CCPA
Analytics	10.0 - 20.0	Weak	SOC-2

Safety-First Workflow

Step 1: Validate Configuration (Always First!)

Never start DP training without validation!

curl -X POST "https://api.synthdata.studio/generators/dp/validate-config" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "your-dataset-id",
    "generator_type": "dp-ctgan",
    "epochs": 50,
    "batch_size": 200,
    "target_epsilon": 10.0
  }'

Successful Validation:

{
  "is_valid": true,
  "errors": [],
  "warnings": ["Batch size is 10% of dataset - good for privacy"],
  "recommended_config": {
    "epochs": 50,
    "batch_size": 200,
    "target_epsilon": 10.0,
    "expected_privacy_level": "Moderate"
  }
}

Common Validation Errors:

{
  "is_valid": false,
  "errors": ["Batch size (500) too large for dataset (1000 rows)"],
  "warnings": [],
  "recommended_config": {
    "batch_size": 100,
    "epochs": 25
  }
}

Step 2: Understand the Warnings

Batch Size Warnings:

• "Batch size is 10% of dataset" = Good for privacy
• "Batch size is 20% of dataset" = Moderate risk
• ❌ "Batch size too large" = High privacy risk

Training Step Warnings:

• "< 500 steps" = Good for privacy
• "500-1000 steps" = Moderate consumption
• ❌ "> 1000 steps" = High privacy budget usage

Generate Privacy-Preserving Data

Healthcare Data Example (HIPAA Compliance)

Scenario

• Dataset: 10,000 patient records with PHI
• Requirements: HIPAA compliance, strong privacy
• Use Case: ML model training for disease prediction

Step 1: Upload and Profile

# Upload healthcare dataset
curl -X POST "https://api.synthdata.studio/datasets/upload" \
  -F "file=@healthcare_data.csv"

# Profile the data
curl -X POST "https://api.synthdata.studio/datasets/{dataset_id}/profile"

# Check for PII
curl -X POST "https://api.synthdata.studio/datasets/{dataset_id}/pii-detection"

Step 2: Validate DP Configuration

curl -X POST "https://api.synthdata.studio/generators/dp/validate-config" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "health-dataset-id",
    "generator_type": "dp-ctgan",
    "epochs": 100,
    "batch_size": 100,
    "target_epsilon": 1.0
  }'

Why these parameters?

• epsilon: 1.0 = Strong HIPAA-compliant privacy
• batch_size: 100 = 1% of 10k dataset (excellent for privacy)
• epochs: 100 = Sufficient training for quality

Step 3: Generate with Privacy

curl -X POST "https://api.synthdata.studio/generators/dataset/{dataset_id}/generate" \
  -H "Content-Type: application/json" \
  -d '{
    "generator_type": "dp-ctgan",
    "num_rows": 10000,
    "target_epsilon": 1.0,
    "epochs": 100,
    "batch_size": 100
  }'

Monitor Progress:

# Check status
curl https://api.synthdata.studio/generators/{generator_id}

Step 4: Review Privacy Report

curl http://localhost:8000/generators/{generator_id}/privacy-report

Expected Report:

{
  "generator_id": "gen-health-123",
  "privacy_level": "Strong",
  "epsilon_achieved": 0.95,
  "epsilon_target": 1.0,
  "budget_utilization": "95%",
  "compliance_frameworks": {
    "HIPAA": {
      "compliant": true,
      "requirements_met": [
        "De-identification of PHI",
        "Privacy Rule compliance",
        "Minimum necessary standard"
      ],
      "recommendations": [
        "Document epsilon value in BAA",
        "Include in HIPAA risk assessment"
      ]
    }
  },
  "safety_assessment": {
    "risk_level": "low",
    "warnings": [],
    "recommendations": [
      "Privacy parameters are well-calibrated for healthcare use"
    ]
  }
}

Financial Data Example (GDPR Compliance)

Scenario

• Dataset: 50,000 customer transaction records
• Requirements: GDPR Article 35 (DPIA), data minimization
• Use Case: Fraud detection model training

Configuration for GDPR

curl -X POST "http://localhost:8000/generators/dp/validate-config" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "finance-dataset-id",
    "generator_type": "dp-tvae",
    "epochs": 50,
    "batch_size": 250,
    "target_epsilon": 5.0
  }'

Why DP-TVAE?

• Faster than DP-CTGAN for large datasets
• Good for mixed numerical/categorical data
• Suitable for transaction data

Generate Financial Data

curl -X POST "http://localhost:8000/generators/dataset/{dataset_id}/generate" \
  -H "Content-Type: application/json" \
  -d '{
    "generator_type": "dp-tvae",
    "num_rows": 50000,
    "target_epsilon": 5.0,
    "epochs": 50,
    "batch_size": 250
  }'

Customer Data Example (CCPA Compliance)

Scenario

• Dataset: 25,000 customer profiles
• Requirements: CCPA right to deletion, data minimization
• Use Case: Customer analytics and personalization

CCPA-Compliant Configuration

curl -X POST "http://localhost:8000/generators/dp/validate-config" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "customer-dataset-id",
    "generator_type": "dp-ctgan",
    "epochs": 75,
    "batch_size": 125,
    "target_epsilon": 10.0
  }'

Compare Privacy vs Quality Trade-offs

Generate Multiple Versions

Let's create three versions with different privacy levels:

Version 1: High Privacy (ε = 1.0)

{
  "generator_type": "dp-ctgan",
  "num_rows": 1000,
  "target_epsilon": 1.0,
  "epochs": 100,
  "batch_size": 50
}

Version 2: Balanced (ε = 5.0)

{
  "generator_type": "dp-ctgan",
  "num_rows": 1000,
  "target_epsilon": 5.0,
  "epochs": 75,
  "batch_size": 100
}

Version 3: High Utility (ε = 10.0)

{
  "generator_type": "ctgan",
  "num_rows": 1000,
  "epochs": 50,
  "batch_size": 200
}

Evaluate All Versions

# Evaluate each generator
for gen_id in "gen-v1" "gen-v2" "gen-v3"; do
  curl -X POST "http://localhost:8000/evaluations/quick/${gen_id}"
done

Compare Results

Version	Epsilon	Quality Score	Privacy Level	Use Case
V1	1.0	0.78	Very Strong	Clinical research
V2	5.0	0.85	Strong	Healthcare analytics
V3	10.0	0.92	Moderate	Customer insights

Key Insights:

• Privacy comes at a cost: Lower ε reduces utility
• Choose based on requirements: Clinical trials need ε < 1.0
• Balance is possible: ε = 5.0 often provides good trade-off

Compliance Documentation

Generate Compliance Reports

# HIPAA compliance report
curl -X POST "http://localhost:8000/generators/{generator_id}/compliance-report" \
  -H "Content-Type: application/json" \
  -d '{"framework": "HIPAA"}'

# GDPR compliance report
curl -X POST "http://localhost:8000/generators/{generator_id}/compliance-report" \
  -H "Content-Type: application/json" \
  -d '{"framework": "GDPR"}'

Model Cards

Create comprehensive documentation:

curl -X POST "http://localhost:8000/generators/{generator_id}/model-card" \
  -H "Content-Type: application/json" \
  -d '{"include_privacy": true}'

Audit Narratives

Generate human-readable audit trails:

curl http://localhost:8000/generators/{generator_id}/audit-narrative

Advanced Privacy Analysis

Membership Inference Testing

Check if synthetic data reveals original records:

curl -X POST "http://localhost:8000/evaluations/run" \
  -H "Content-Type: application/json" \
  -d '{
    "generator_id": "gen-123",
    "dataset_id": "original-dataset-id",
    "include_privacy": true
  }'

Good Privacy Result:

{
  "privacy_tests": {
    "membership_inference": {
      "attack_success_rate": 0.51,
      "baseline_accuracy": 0.5,
      "privacy_score": 0.98,
      "risk_level": "low",
      "interpretation": "No significant membership inference risk"
    }
  }
}

Attribute Inference Testing

Test for sensitive attribute disclosure:

{
  "privacy_tests": {
    "attribute_inference": {
      "target_attribute": "salary",
      "attack_accuracy": 0.12,
      "baseline_accuracy": 0.1,
      "privacy_score": 0.8,
      "risk_level": "low"
    }
  }
}

Parameter Optimization

Systematic Parameter Search

Start Conservative

{
  "target_epsilon": 1.0,
  "epochs": 50,
  "batch_size": 100
}

Test Quality Impact

• Run evaluation after each change
• Monitor statistical similarity scores
• Check ML utility metrics

Adjust Based on Results

// If quality too low, increase epsilon
{
  "target_epsilon": 2.0,
  "epochs": 75,
  "batch_size": 100
}

// If privacy too weak, decrease epsilon
{
  "target_epsilon": 0.5,
  "epochs": 100,
  "batch_size": 50
}

Automated Recommendations

Use the API to get parameter suggestions:

curl "http://localhost:8000/generators/dp/recommended-config?dataset_id={id}&desired_quality=balanced"

Common Privacy Issues

"Epsilon Too Low" Error

Problem: Target epsilon cannot be achieved with current parameters

Solutions:

1. Increase batch size
2. Reduce epochs
3. Use DP-TVAE instead of DP-CTGAN
4. Accept higher epsilon value

Poor Quality with Strong Privacy

Problem: ε = 0.1 produces unusable synthetic data

Solutions:

1. Assess if ε = 0.1 is truly required
2. Consider ε = 1.0 for better utility
3. Use domain-specific privacy approaches
4. Implement hybrid approaches (DP + other techniques)

Membership Inference Failures

Problem: Privacy attacks succeed too often

Solutions:

1. Reduce epsilon further
2. Increase noise multiplier
3. Use different DP algorithm
4. Implement additional privacy techniques

Compliance Checklist

HIPAA Compliance

• ε ≤ 1.0 for PHI data
• Document privacy parameters in BAA
• Include in HIPAA risk assessment
• Regular privacy audits
• Staff training on DP concepts

GDPR Compliance

• Data Protection Impact Assessment (DPIA)
• Privacy by design documentation
• Lawful processing justification
• Data minimization evidence
• Individual rights implementation

CCPA Compliance

• Privacy notice updates
• Data usage documentation
• Individual rights responses
• Service provider agreements
• Security assessments

Tutorial Complete!

What You Accomplished

Understood differential privacy concepts Validated DP configurations safely Generated privacy-preserving synthetic data Reviewed compliance documentation Compared privacy-utility trade-offs Created audit-ready reports

Your Privacy-Compliant Dataset

You now have:

• Mathematically private synthetic data
• Compliance documentation for HIPAA/GDPR/CCPA
• Privacy guarantees with known epsilon bounds
• Audit trails for regulatory reviews

Advanced Topics

Next Steps

1. Quality Evaluation Tutorial: Deep dive into evaluation metrics
2. Compliance Reporting Tutorial: Generate audit documentation
3. API Integration Guide: Build custom integrations

Advanced Privacy Techniques

• Hybrid Approaches: Combine DP with other privacy methods
• Local Differential Privacy: Client-side privacy
• Federated Learning: Privacy-preserving distributed training
• Homomorphic Encryption: Compute on encrypted data

Research Directions

• Dynamic Epsilon: Adjust privacy based on data sensitivity
• Attribute-Specific Privacy: Different epsilon for different columns
• Temporal Privacy: Privacy over time series data
• Composition Attacks: Advanced privacy threat models

Resources

Documentation

• Privacy Features Guide: Complete DP reference
• API Examples: Code examples and API usage
• Compliance Guide: Regulatory requirements

Research Papers

• "Deep Learning with Differential Privacy" (Abadi et al.)
• "Differential Privacy: A Survey" (Dwork & Roth)
• "Privacy-Preserving Synthetic Data" (Bowen & Liu)

Tools & Libraries

• Opacus: PyTorch differential privacy library
• TensorFlow Privacy: TF differential privacy
• PySyft: Privacy-preserving machine learning

---

Congratulations! You now know how to generate mathematically private synthetic data. Your datasets are ready for HIPAA, GDPR, and CCPA compliance!