Privacy Features & Differential Privacy

Learn how to use Synthetic Data Studio's differential privacy features to generate mathematically private synthetic data with regulatory compliance.

Differential Privacy Fundamentals

What is Differential Privacy?

Differential Privacy (DP) is a mathematical framework that provides strong privacy guarantees for data analysis. It ensures that the output of any analysis is statistically indistinguishable whether or not any individual's data is included.

Key Concept: A DP algorithm adds carefully calibrated noise to ensure that individual records cannot be distinguished in the output.

Privacy Parameters

Epsilon (ε) - Privacy Budget

• Definition: Maximum privacy loss per individual
• Range: 0.1 (very private) to 100+ (less private)
• Interpretation: Lower ε = stronger privacy guarantees
• Typical Values: 0.1-10.0 for production use

Delta (δ) - Failure Probability

• Definition: Probability that privacy guarantee fails
• Typical Value: 1/n (where n = dataset size)
• Auto-calculated in Synthetic Data Studio

Privacy Levels

Epsilon (ε)	Privacy Level	Use Case	Compliance
< 1.0	Very Strong	Clinical trials, genomic data	HIPAA (PHI)
1-5	Strong	Healthcare, financial records	GDPR Article 9
5-10	Moderate	Customer data, HR records	CCPA
10-20	Weak	Aggregated analytics	SOC-2
> 20	Minimal	Non-sensitive data	None required

Safety System

3-Layer Privacy Protection

Synthetic Data Studio implements a comprehensive safety system:

1. Pre-Training Validation

Purpose: Prevent catastrophic privacy failures before training starts

# Validate DP configuration
curl -X POST "http://localhost:8000/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
  }'

Validation Checks:

• Sampling rate validation (< 10% recommended)
• Training steps calculation
• Noise multiplier computation
• Privacy budget verification

2. Runtime Privacy Accounting

Purpose: Track actual privacy expenditure during training

• Rényi Differential Privacy (RDP) accounting
• Composition theorems for multiple training steps
• Real-time budget tracking
• Automatic noise calibration

3. Post-Training Verification

Purpose: Confirm final privacy guarantees

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

Verification Includes:

• Actual epsilon achieved
• Privacy budget utilization
• Safety warnings and recommendations
• Compliance framework mapping

Using Differential Privacy

Step 1: Validate Configuration

Always validate before training:

curl -X POST "http://localhost:8000/generators/dp/validate-config" \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "550e8400-e29b-41d4-a716-446655440000",
    "generator_type": "dp-ctgan",
    "epochs": 50,
    "batch_size": 200,
    "target_epsilon": 10.0
  }'

Successful Response:

{
  "is_valid": true,
  "errors": [],
  "warnings": [
    "Batch size (200) is 10% of dataset size - good for privacy"
  ],
  "requested_config": {
    "epochs": 50,
    "batch_size": 200,
    "target_epsilon": 10.0,
    "sampling_rate": "10.00%"
  },
  "recommended_config": {
    "epochs": 50,
    "batch_size": 200,
    "target_epsilon": 10.0,
    "expected_privacy_level": "Moderate"
  }
}

Step 2: Get Recommendations

Let the system suggest optimal parameters:

curl "http://localhost:8000/generators/dp/recommended-config?dataset_id=550e8400-e29b-41d4-a716-446655440000&desired_quality=balanced"

Quality Options:

• high_privacy: ε < 5, prioritizes privacy over quality
• balanced: ε ≈ 10, good balance of privacy and utility
• high_quality: ε ≈ 15, prioritizes quality over privacy

Step 3: Generate with Privacy

curl -X POST "http://localhost:8000/generators/dataset/550e8400-e29b-41d4-a716-446655440000/generate" \
  -H "Content-Type: application/json" \
  -d '{
    "generator_type": "dp-ctgan",
    "num_rows": 1000,
    "target_epsilon": 10.0,
    "epochs": 50,
    "batch_size": 200
  }'

Step 4: Review Privacy Report

After generation, check the privacy report:

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

Sample Report:

{
  "generator_id": "gen-123",
  "privacy_level": "Moderate",
  "epsilon_achieved": 9.8,
  "epsilon_target": 10.0,
  "budget_utilization": "98%",
  "compliance_frameworks": {
    "HIPAA": {
      "compliant": true,
      "requirements_met": ["De-identification", "Privacy Rule"],
      "recommendations": ["Document privacy parameters in BAA"]
    },
    "GDPR": {
      "compliant": true,
      "requirements_met": ["Article 5", "Article 25"],
      "recommendations": ["Include in DPIA documentation"]
    }
  },
  "safety_assessment": {
    "risk_level": "low",
    "warnings": [],
    "recommendations": [
      "Privacy parameters are well-calibrated",
      "Consider ε=5.0 for higher privacy if needed"
    ]
  }
}

DP Method Comparison

DP-CTGAN vs DP-TVAE

Feature	DP-CTGAN	DP-TVAE
Quality	Excellent	Good
Speed	Slower	2-3x faster
Memory	Higher	Lower
Best For	Complex data	Mixed types
Training Time	10-30 min	5-15 min

When to Use Each Method

Choose DP-CTGAN when:

• Maximum data quality is required
• Complex correlations need preservation
• Training time is not a constraint
• Dataset has intricate patterns

Choose DP-TVAE when:

• Faster generation is needed
• Dataset has mixed data types
• Memory constraints exist
• Prototyping or iteration speed matters

⚙ Advanced Configuration

Fine-Tuning Privacy Parameters

{
  "target_epsilon": 5.0,
  "target_delta": 0.0001,
  "max_grad_norm": 1.0,
  "noise_multiplier": "auto",
  "epochs": 100,
  "batch_size": 100
}

Understanding Noise Multiplier

The noise multiplier controls privacy-utility trade-off:

• Lower values (< 0.5): Less noise, better utility, weaker privacy
• Medium values (0.5-2.0): Balanced noise, good privacy-utility trade-off
• Higher values (> 2.0): More noise, stronger privacy, reduced utility

Custom Delta Values

{
  "target_delta": 1e-5,  // Instead of auto-calculated 1/n
  "target_epsilon": 1.0
}

Note: Lower delta values require more noise, reducing data quality.

Compliance Frameworks

HIPAA Compliance

Requirements Met:

• De-identification of Protected Health Information (PHI)
• Safe Harbor method compliance
• Expert determination documentation
• Privacy Rule compliance

Implementation:

{
  "target_epsilon": 1.0,
  "generator_type": "dp-ctgan",
  "compliance_framework": "HIPAA"
}

GDPR Compliance

Requirements Met:

• Data minimization (Article 5)
• Privacy by design (Article 25)
• Lawful processing (Article 6)
• Data protection impact assessment (Article 35)

Implementation:

{
  "target_epsilon": 5.0,
  "generator_type": "dp-tvae",
  "compliance_framework": "GDPR"
}

CCPA Compliance

Requirements Met:

• Right to know about data collection
• Right to delete personal information
• Right to opt-out of data sales
• Data minimization requirements

SOC-2 Compliance

Requirements Met:

• Security principle
• Availability principle
• Processing integrity
• Confidentiality principle

Privacy Evaluation

Membership Inference Testing

Check if synthetic data could reveal 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
  }'

Attribute Inference Testing

Test for potential attribute disclosure:

{
  "privacy_tests": {
    "membership_inference": true,
    "attribute_inference": true,
    "target_attributes": ["income", "age", "location"]
  }
}

Best Practices

Privacy-First Development

1. Always Validate First: Use /dp/validate-config before training
2. Start Conservative: Begin with lower epsilon values
3. Monitor Budget: Track privacy expenditure throughout training
4. Document Decisions: Record privacy parameters for compliance

Parameter Selection Guidelines

For Healthcare Data (HIPAA)

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

For Financial Data (GDPR)

{
  "target_epsilon": 5.0,
  "generator_type": "dp-tvae",
  "epochs": 50,
  "batch_size": 100
}

For General Business Data

{
  "target_epsilon": 10.0,
  "generator_type": "dp-ctgan",
  "epochs": 50,
  "batch_size": 200
}

Quality vs Privacy Trade-offs

High Privacy (ε < 5):

• Strong privacy guarantees
• Reduced data quality
• More noise in results
• Regulatory compliance

Balanced Approach (ε = 5-10):

• Good privacy protection
• Reasonable data quality
• Practical for most use cases
• Compliance-friendly

High Quality (ε > 10):

• Weaker privacy guarantees
• Better data utility
• Less noise in results
• May not meet strict compliance

� Common Issues & Solutions

"Sampling Rate Too High"

Error: Batch size too large (>20% of dataset)

Solution:

{
  "batch_size": 100,  // Reduce from 500
  "epochs": 25        // May need to increase epochs
}

"Privacy Budget Exceeded"

Error: Target epsilon too low for training parameters

Solutions:

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

"Training Unstable"

Error: Privacy accounting failed - training unstable

Solutions:

1. Reduce learning rate
2. Increase max_grad_norm
3. Use smaller batch sizes
4. Validate configuration first

"Poor Quality Results"

Issue: Synthetic data quality is low despite DP guarantees

Solutions:

1. Increase epochs gradually
2. Use DP-CTGAN over DP-TVAE
3. Adjust noise multiplier
4. Consider non-DP methods if privacy requirements allow

Monitoring & Auditing

Privacy Budget Tracking

Monitor privacy expenditure:

# Check generator status during training
curl http://localhost:8000/generators/{generator_id}

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

Audit Trail

Maintain records for compliance:

• Configuration parameters used
• Validation results
• Privacy reports generated
• Quality evaluation results
• Compliance framework mappings

Performance Monitoring

Track DP training metrics:

• Training time vs privacy parameters
• Memory usage patterns
• Convergence behavior
• Quality metrics over time

Integration with Compliance

Automated Compliance Reporting

Generate compliance documentation:

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

Model Cards

Create comprehensive model 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

Next Steps

After implementing privacy features:

1. Evaluate Quality - Assess privacy-utility trade-offs
2. Generate Compliance Reports - Create audit documentation
3. Use AI Features - Get AI-powered privacy insights

---

Need help with privacy parameters? Use the /dp/recommended-config endpoint or check our Troubleshooting Guide.