2nd/10_Wiki/Topics/AI_and_ML/Data-Ethics and Privacy.md

---
id: wiki-2026-0508-data-ethics-privacy
title: Data Ethics and Privacy
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [data ethics, privacy, GDPR, CCPA, differential privacy, federated learning, k-anonymity, PII]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [ethics, privacy, gdpr, ccpa, differential-privacy, federated-learning, ai-governance, pii]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: ethics / law / cryptography
  applicable_to: [Data Governance, ML Privacy, Compliance]
---

# Data Ethics and Privacy

## 매 한 줄
> **"매 can we의 X — 매 should we"**. 매 GDPR (EU 2018) + CCPA (CA) + 매 EU AI Act (2024). 매 modern technique: differential privacy, federated learning, homomorphic encryption, ZK proof. 매 LLM 시대 의 training data + memorization 의 new challenge.

## 매 핵심 principle

### 매 Fair Information Practice Principles (FIPPs)
1. **Notice / Transparency**.
2. **Consent / Choice**.
3. **Access / Participation**.
4. **Integrity / Security**.
5. **Enforcement / Redress**.

### GDPR (EU, 2018)
- **Lawful basis**: 매 6 (consent, contract, legal, vital, public, legitimate interest).
- **Data minimization**.
- **Purpose limitation**.
- **Right to access / rectify / erasure / portability / object**.
- **DPIA** (Data Protection Impact Assessment).
- **DPO** (Data Protection Officer).
- 매 fine: 매 4% global revenue.

### CCPA / CPRA (California)
- 매 sell 의 opt-out.
- 매 sensitive data 의 limit.

### 매 EU AI Act (2024)
- 매 risk tier (unacceptable / high / limited / minimal).
- 매 high-risk: 매 audit, 매 human oversight.

### 매 privacy techniques

#### Anonymization (irreversible)
- 매 PII 의 remove + 매 generalize.
- 매 k-anonymity (Sweeney 2002).
- 매 l-diversity, t-closeness.

#### Pseudonymization (reversible with key)
- 매 PII 의 hash / token.
- 매 GDPR 의 lighter requirement.

#### Differential Privacy (Dwork 2006)
- 매 noise injection.
- 매 mathematical guarantee (ε).
- 매 Apple, Google, US Census.

#### Federated Learning
- 매 raw data 의 leave 의 X.
- 매 model update 만.
- 매 mobile, healthcare.

#### Homomorphic Encryption
- 매 encrypted 의 compute.
- 매 expensive (still).

#### Secure Multi-party Computation (MPC)
- 매 multiple party 의 compute 의 share X.

#### Zero-Knowledge Proof
- 매 prove without reveal.
- 매 ZK-ML 의 emerging.

### LLM-specific privacy
- **Training data leak**: 매 verbatim memorization.
- **Membership inference**: 매 was 매 X 의 train data?
- **Model inversion**: 매 model 의 input 의 reconstruct.
- **Mitigation**: dedup, DP-SGD, scrubbing, prompt safety.

### 매 design pattern
1. **Privacy by Design** (Cavoukian).
2. **Data minimization**.
3. **Encryption at rest + in transit**.
4. **Access control + audit log**.
5. **Retention policy**.
6. **Right to erasure pipeline**.

## 💻 패턴

### PII detection + redaction
```python
import re
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine

analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()

def anonymize(text, language='en'):
    results = analyzer.analyze(text=text, language=language)
    return anonymizer.anonymize(text=text, analyzer_results=results).text

# 매 example
print(anonymize('John Smith, ssn 123-45-6789, lives at 123 Main St.'))
# <PERSON>, ssn <US_SSN>, lives at <LOCATION>.
```

### k-Anonymity
```python
def k_anonymize(df, quasi_identifiers, k=5):
    """매 매 group 의 size ≥ k."""
    grouped = df.groupby(quasi_identifiers).size()
    valid_groups = grouped[grouped >= k].index
    return df[df.set_index(quasi_identifiers).index.isin(valid_groups)]

# 매 generalize age, zip → 매 group ≥ 5
```

### Differential Privacy (Laplace mechanism)
```python
import numpy as np

def laplace_mechanism(query_result, sensitivity, epsilon):
    """매 ε-differential privacy."""
    noise = np.random.laplace(0, sensitivity / epsilon)
    return query_result + noise

# 매 example: 매 count of users with X
count_true = df['has_x'].sum()
count_dp = laplace_mechanism(count_true, sensitivity=1, epsilon=1.0)
```

### DP-SGD (training)
```python
from opacus import PrivacyEngine

privacy_engine = PrivacyEngine()
model, optimizer, train_loader = privacy_engine.make_private(
    module=model,
    optimizer=optimizer,
    data_loader=train_loader,
    noise_multiplier=1.0,
    max_grad_norm=1.0,
)

# 매 training 의 DP guarantee
for x, y in train_loader:
    optimizer.zero_grad()
    loss = F.cross_entropy(model(x), y)
    loss.backward()
    optimizer.step()

epsilon = privacy_engine.get_epsilon(delta=1e-5)
print(f'(ε={epsilon}, δ={1e-5})-DP')
```

### Federated Learning (Flower)
```python
import flwr as fl

class Client(fl.client.NumPyClient):
    def get_parameters(self, config):
        return [val.cpu().numpy() for val in model.state_dict().values()]
    
    def fit(self, parameters, config):
        # 매 local training (raw data 의 stay local)
        set_parameters(model, parameters)
        train(model, local_dataset)
        return get_parameters(model), len(local_dataset), {}
    
    def evaluate(self, parameters, config):
        loss, accuracy = test(model, local_dataset)
        return loss, len(local_dataset), {'accuracy': accuracy}

fl.client.start_client(server_address='central:8080', client=Client())
```

### Right to Erasure (GDPR)
```python
def erase_user(user_id):
    """매 GDPR Art. 17."""
    # 매 1. primary
    db.users.delete(user_id)
    
    # 매 2. derived (cache, analytics, ML training)
    cache.invalidate(f'user:{user_id}')
    analytics_db.delete_user_events(user_id)
    
    # 매 3. backup (delayed)
    schedule_backup_purge(user_id, after_days=30)
    
    # 매 4. ML model — 매 retrain or 매 forget request
    if user_in_training_data(user_id):
        machine_unlearning(user_id) or schedule_retrain()
    
    # 매 5. log + audit
    audit_log.write({'action': 'erase', 'user_id': user_id, 'date': now()})
```

### Access control + audit
```python
def access_pii(user_id, requester, reason):
    if not requester.has_role('data_steward'):
        raise PermissionError()
    
    audit_log.write({
        'requester': requester.id,
        'subject': user_id,
        'reason': reason,
        'timestamp': datetime.now(),
        'data_accessed': ['email', 'phone'],
    })
    
    return db.users.get(user_id, fields=['email', 'phone'])
```

### LLM verbatim leak detection
```python
def check_training_data_leak(model, training_corpus, n_samples=100):
    """매 model 의 verbatim 의 reproduce?"""
    leaks = []
    for chunk in random.sample(training_corpus, n_samples):
        prefix = chunk[:100]
        completion = model.generate(prefix, max_tokens=200)
        if chunk[100:300] in completion:
            leaks.append({'chunk': chunk[:50] + '...', 'leaked': True})
    return leaks
```

### Membership inference attack (defense check)
```python
def membership_inference_test(model, member_data, non_member_data):
    """매 model 의 매 specific 의 in train data 의 distinguish?"""
    member_loss = [loss(model, x, y) for x, y in member_data]
    non_member_loss = [loss(model, x, y) for x, y in non_member_data]
    
    # 매 simple threshold attack
    threshold = np.median(member_loss + non_member_loss)
    
    correct = sum(1 for l in member_loss if l < threshold) + \
              sum(1 for l in non_member_loss if l > threshold)
    accuracy = correct / (len(member_loss) + len(non_member_loss))
    
    if accuracy > 0.55:
        return 'WARN: membership inference vulnerability'
    return 'OK'
```

### Data retention policy
```python
def retention_purge():
    """매 매 day 의 cron."""
    # 매 GDPR + business 의 retention.
    db.events.delete(where=f"created_at < NOW() - INTERVAL '7 years'")
    db.user_logs.delete(where=f"created_at < NOW() - INTERVAL '90 days'")
    db.analytics_raw.delete(where=f"created_at < NOW() - INTERVAL '13 months'")
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| EU users | GDPR compliance (DPO + DPIA + erasure) |
| US users | CCPA + sector law (HIPAA, FERPA) |
| Sensitive aggregate stats | Differential Privacy |
| Cross-org training | Federated Learning |
| Encrypted compute | Homomorphic / MPC |
| LLM training | DP-SGD + dedup + scrub |
| Identity-needed | Pseudonymize + access control |
| Public data | Anonymize (k-anonymity) |

**기본값**: Privacy by Design + 매 access control + 매 retention + 매 erasure pipeline.

## 🔗 Graph
- 부모: [[AI-Ethics]] · [[Privacy]]
- 변형: [[GDPR]] · [[CCPA]] · [[Differential-Privacy]] · [[Federated-Learning]] · [[Homomorphic-Encryption]]
- 응용: [[k-Anonymity]]
- Adjacent: [[Algorithmic-Fairness]] · [[Authenticity]] · [[AI-Sovereignty]] · [[Data-Flywheel-Effect]] · [[Anthropomorphism]]

## 🤖 LLM 활용
**언제**: 매 product privacy review. 매 GDPR audit. 매 ML privacy. 매 cross-border data transfer.
**언제 X**: 매 specific legal advice (lawyer). 매 medical clinical (HIPAA expert).

## ❌ 안티패턴
- **Anonymization 의 false sense** (linkage attack 가능): 매 k-anonymity + 매 l-diversity.
- **Long retention without business need**: 매 GDPR violation.
- **No erasure pipeline**: 매 right 의 fulfill X.
- **PII in logs**: 매 invisible leak.
- **Federated 의 raw data 의 leak** (model invert): 매 DP-SGD 도 필요.
- **Consent 의 invalid** (forced, vague).

## 🧪 검증 / 중복
- Verified (GDPR text, Apple DP, Dwork DP paper, Sweeney k-anonymity).
- 신뢰도 A.
- Related: [[Algorithmic-Fairness]] · [[Authenticity]] · [[Bias-Correction-Algorithm]] · [[Anthropomorphism]] · [[Atmospheric-Intelligence]] (privacy challenges).

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — GDPR + DP + FL + 매 Presidio / k-anon / Opacus / Flower / erasure code |