ML Training Data: Reproducible Privacy Presets

ML Training Data: Reproducible Privacy Presets

Ang Problema: Non-Reproducible Anonymization

Ang karamihan ng ML teams ay nag-anonymize ng training data using non-deterministic methods:

import random
import hashlib

def anonymize_for_ml(data):
    """Non-reproducible anonymization"""
    result = []
    for record in data:
        # Using random salt - different bawat run
        salt = os.urandom(16)
        anonymized_name = hashlib.sha256((record['name'] + salt).encode()).hexdigest()
        
        result.append({
            'name_hash': anonymized_name,  # Different output each time!
            'email': record['email'],      # Not anonymized at all!
        })
    return result

# Run 1
dataset_v1 = anonymize_for_ml(raw_data)

# Run 2 (same input, different output)
dataset_v2 = anonymize_for_ml(raw_data)

# dataset_v1 != dataset_v2 (due to random salt)
# Model trained on v1 and v2 will have different hashes

Problem:

• ❌ Cannot reproduce same training dataset
• ❌ Cannot link records across runs (for validation)
• ❌ Cannot audit which anonymization was used
• ❌ Not GDPR-compliant (no deterministic trail)

Ang Solution: Deterministic Reproducible Presets

Ang reproducible anonymization uses deterministic hashing without random salt:

Hakbang 1: Define Reproducible Preset

{
  "preset_name": "ML Training Data - GDPR Safe",
  "version": "1.0",
  "reproducibility": "DETERMINISTIC",
  "seed": "organization_secret_key_2025",
  "rules": [
    {
      "entity_type": "PERSON",
      "operator": "hash",
      "algorithm": "HMAC-SHA256",
      "secret_key": "organization_secret_key_2025",
      "deterministic": true
    },
    {
      "entity_type": "EMAIL_ADDRESS",
      "operator": "hash",
      "algorithm": "SHA-256",
      "no_salt": true,
      "deterministic": true
    },
    {
      "entity_type": "PHONE_NUMBER",
      "operator": "mask",
      "retain_format": true,
      "last_digits": 4,
      "deterministic": true
    }
  ]
}

Hakbang 2: Implement Reproducible Anonymization

import hashlib
import hmac

def anonymize_reproducible(data, preset):
    """Deterministic, reproducible anonymization"""
    result = []
    
    secret_key = preset['seed']
    
    for record in data:
        anonymized_name = hmac.new(
            key=secret_key.encode(),
            msg=record['name'].encode(),
            digestmod=hashlib.sha256
        ).hexdigest()
        
        anonymized_email = hashlib.sha256(
            record['email'].encode()
        ).hexdigest()
        
        result.append({
            'person_hash': anonymized_name,
            'email_hash': anonymized_email,
            'phone_masked': record['phone'][-4:]
        })
    
    return result

# Run 1
preset = load_preset("ML Training Data - GDPR Safe")
dataset_v1 = anonymize_reproducible(raw_data, preset)

# Run 2 (same input, same output)
dataset_v2 = anonymize_reproducible(raw_data, preset)

# dataset_v1 == dataset_v2 ✅ (deterministic)
# Hash of "John Smith" is always the same

Hakbang 3: Version Control Dataset

# Generate training dataset v1
python anonymize.py --preset "ML Training Data - GDPR Safe" --output dataset-v1.csv

# Compute checksum
sha256sum dataset-v1.csv > dataset-v1.sha256

# Store in version control
git add dataset-v1.csv dataset-v1.sha256
git commit -m "ML training data v1 - GDPR safe anonymization"

# Later, regenerate same dataset
python anonymize.py --preset "ML Training Data - GDPR Safe" --output dataset-v1-verify.csv
sha256sum dataset-v1-verify.csv

# Verify checksums match
# ✅ Same dataset reproduced exactly

Ang Benefits ng Reproducible Presets

Ang GDPR Compliance Use Case

Ang GDPR Article 32 requires documented security measures. With reproducible presets:

DPA Audit Question: "Can you prove that this training dataset was anonymized according to your policy?"

Without Reproducibility:

"We anonymized it... but we can't reproduce it." Result: ❌ FAILING

With Reproducibility:

"Here's the preset, the seed, and the version checksum. We can regenerate the exact same dataset." Result: ✅ PASSING

Real-World Case: Healthcare ML Company

Isang healthcare ML company ay nag-develop ng model para sa disease prediction:

Problem:

- Trained model A sa anonymized dataset v1
- Tried to retrain model B sa same dataset
- Output ay iba dahil random salt sa anonymization
- Cannot reproduce exactly what data was used
- DPA audit fails: "Cannot verify dataset anonymization"

Solution: Reproducible Preset

- Used deterministic HMAC-SHA256 preset
- Trained model A sa anonymized dataset v1
- Retraining model B produced identical anonymization
- Checksum verification: ✅ exact match
- DPA audit passes: "Here's the preset, seed, and verification"

Benefits:
- ✅ Reproducible training data
- ✅ GDPR compliant with audit trail
- ✅ Can validate model versions
- ✅ Can detect dataset drift

Ang Best Practice

1. Use deterministic hashing - no random salts
2. Store secret key securely - in secrets manager
3. Version control presets - document changes
4. Compute dataset checksums - for verification
5. Audit trail - log each dataset generation

Ang reproducible anonymization presets ay essential para sa ML teams na may GDPR compliance requirements.