Skip to main content

10.3 Data Privacy & Compliance

AI-Generated Content

AI-generated content may contain errors. Always verify against official sources.

10.3 Data Privacy & Compliance

Key Concepts: On-premise vs cloud · PII handling · Data residency (UAE/GDPR)

Official Docs: OpenAI Data Privacy · Azure OpenAI Data Residency

Key Regulations to Know​

RegulationRegionKey requirement
GDPREU/EEAData must not leave EU without adequacy; right to erasure
UAE PDPLUAEPersonal data of UAE residents protected; consent required
HIPAAUSAHealth data must be encrypted and audit-logged
PDPAThailand/SingaporeExplicit consent for personal data processing

Decision: Cloud API vs On-Premise​

Does your data include PII or regulated data?
         │
   No    │  Yes
═════════╫════════
↓         ↓
Cloud    Does your cloud provider offer data residency?
 API        │
      No    │  Yes
══════════╫═══════
↓           ↓
On-premise  Cloud with data residency
(Ollama/    (Azure OpenAI EU region)
vLLM)

PII Detection and Redaction​

Before sending data to any cloud LLM, detect and redact PII:

import re
from openai import OpenAI

client = OpenAI()

# Simple rule-based PII redaction
def redact_pii(text: str) -> tuple[str, dict]:
    """Redact PII from text. Returns redacted text and a mapping for restoration."""
    redacted = text
    mapping = {}
    counter = {"email": 0, "phone": 0, "id": 0}
    
    # Email addresses
    emails = re.findall(r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}', text)
    for email in emails:
        placeholder = f"[EMAIL_{counter['email']}]"
        mapping[placeholder] = email
        redacted = redacted.replace(email, placeholder)
        counter['email'] += 1
    
    # Phone numbers (international format)
    phones = re.findall(r'\+?[1-9]\d{7,14}', text)
    for phone in phones:
        placeholder = f"[PHONE_{counter['phone']}]"
        mapping[placeholder] = phone
        redacted = redacted.replace(phone, placeholder)
        counter['phone'] += 1
    
    # UAE Emirates IDs (784-XXXX-XXXXXXX-X)
    uae_ids = re.findall(r'784-\d{4}-\d{7}-\d', text)
    for id_ in uae_ids:
        placeholder = f"[UAE_ID_{counter['id']}]"
        mapping[placeholder] = id_
        redacted = redacted.replace(id_, placeholder)
        counter['id'] += 1
    
    return redacted, mapping

def restore_pii(text: str, mapping: dict) -> str:
    """Restore redacted PII after LLM processing."""
    for placeholder, original in mapping.items():
        text = text.replace(placeholder, original)
    return text

# Example
user_message = "My email is [email protected] and my phone is +971501234567. Please analyse my contract."
redacted, mapping = redact_pii(user_message)
print(f"Redacted: {redacted}")

# Safe to send to cloud API
resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": redacted}],
)
llm_response = resp.choices[0].message.content

# Restore if needed
final = restore_pii(llm_response, mapping)

Audit Logging​

For regulated industries, every LLM call must be logged:

import json
import hashlib
from datetime import datetime
from pathlib import Path

def audit_log(user_id: str, prompt_hash: str, response_hash: str, tokens: int, model: str):
    """Log LLM calls without storing actual content (for GDPR compliance)."""
    entry = {
        "timestamp": datetime.utcnow().isoformat(),
        "user_id": user_id,
        "prompt_sha256": prompt_hash,      # Hash, not raw text
        "response_sha256": response_hash,  # Hash, not raw text
        "tokens_used": tokens,
        "model": model,
    }
    with open("audit.jsonl", "a") as f:
        f.write(json.dumps(entry) + "\n")

# Usage
from hashlib import sha256
audit_log(
    user_id="user_abc123",
    prompt_hash=sha256("user message".encode()).hexdigest(),
    response_hash=sha256("llm response".encode()).hexdigest(),
    tokens=150,
    model="gpt-4o-mini",
)

Common Mistakes​

Common Mistakes
  1. Sending raw medical/legal data to cloud APIs — unless you have a Business Associate Agreement (BAA) with the provider, this may violate HIPAA/GDPR.
  2. Rule-based PII detection only — simple regex misses names, addresses, custom ID formats. Consider using a dedicated PII detection model (e.g., spaCy NER, Microsoft Presidio).
  3. Logging full prompts — storing raw prompts containing user data creates a compliance liability. Log hashes or metadata only.
  4. Forgetting about RAG data — if your vector database contains personal data, retrieval can expose that data to the LLM and downstream systems.

Quick Quiz​

Test Your Understanding

Q1. What does data residency mean?
A1. Ensuring that data is stored and processed only within a specified geographic region — e.g., EU data stays in EU servers to comply with GDPR.

Q2. Why should you log prompt hashes rather than raw prompts?
A2. Raw prompts may contain personal data, creating a GDPR liability. Hashes allow integrity verification without storing sensitive content.

Q3. What is Microsoft Presidio?
A3. An open-source PII detection and anonymisation framework that supports 20+ entity types and multiple languages, more comprehensive than regex.

Student Exercise​

Exercise 11.3 — PII redaction pipeline
Extend the redact_pii function to also detect: names (using spacy NER), IBAN numbers, and passport numbers. Build a test with 10 examples. Measure false positive and false negative rates.

Further Reading​

Next → 11.4 Monitoring & Observability