Privacy by Design

Generative AI applications are literal data vacuums. They ingest PDFs, chat logs, and database records. If you are not careful, your LLM might start telling one customer about the Social Security Number of another.

In the AWS Certified Generative AI Developer – Professional exam, Domain 1 and Domain 3 overlap heavily here. You must demonstrate that you can build a "Privacy-First" architecture that identifies and protects Personally Identifiable Information (PII) and Protected Health Information (PHI).

---

1. What is Sensitive Data in GenAI?

Data sensitivity in the context of AI is divided into three levels:

1. Public: Marketing materials, public bios.
2. Internal/Confidential: Strategic plans, internal Slack logs.
3. Highly Sensitive (PII/PHI): Name, Address, SSN, Credit Card numbers, Medical history.

The Professional Goal: Sensitive data should never be stored in your vector database or sent to a Foundation Model unless it is absolutely necessary and heavily encrypted.

---

2. Tools for Data Identification

Before you can protect data, you must find it.

Amazon Macie

Automatically discovers and protects sensitive data at scale in S3. It uses Machine Learning to find SSNs, API keys, and passport numbers hidden in your AI data lake.

Amazon Comprehend (PII Detection)

A natural language processing (NLP) service that can identify PII in raw text. It is perfect for a pre-processing step in your ETL pipeline.

graph LR
    S3[S3 Raw Bucket] -->|Event| L[AWS Lambda]
    L -->|Detect PII| C[Amazon Comprehend]
    C -->|Identify SSN| M[Mask/Redact Logic]
    M -->|Anonymized Text| S3C[S3 Clean Bucket]
    S3C --> KB[Bedrock Knowledge Base]

---

3. Masking and Redaction Strategies

When you find sensitive data, you have three choices:

Strategy	Action	Use Case
Redaction	Replace with [REDACTED].	General analysis where the specific person doesn't matter.
Masking	Replace with a generic label like [NAME_1].	Tasks where the "existence" of a relationship matters but the identity doesn't.
Hashing	Replace with a unique code (e.g., SHA-256).	Tasks where you need to track the same person across multiple documents without knowing who they are.

---

4. Amazon Bedrock Guardrails (PII Filtering)

AWS provides a managed solution for this within Amazon Bedrock. You can configure a Guardrail that automatically:

1. Detects PII in the user's prompt (Input).
2. Detects PII in the model's response (Output).
3. Blocks or masks the PII based on your rules.

Scenario: A customer types their credit card number into your chatbot. Pro Action: A Bedrock Guardrail detects the pattern XXXX-XXXX-XXXX-XXXX and masks it before it ever reaches the LLM, protecting the model provider from receiving sensitive data.

---

5. Data Residency and Isolation

Compliance often demands that data never leaves a specific geographic border.

• VPC Endpoints: Ensure traffic between your app and Bedrock never touches the public internet.
• Region Pining: If you use eu-central-1 (Frankfurt), the compute and the inference stay within the EU.
• Zero Data Training: AWS explicitly states that customer data is not used to train or improve the base models in Amazon Bedrock. This is a critical point for enterprise compliance.

---

6. Code Example: PII Detection with Boto3 (Comprehend)

import boto3

comprehend = boto3.client('comprehend', region_name='us-east-1')

def mask_pii(text):
    # Detect PII entities
    response = comprehend.detect_pii_entities(
        Text=text,
        LanguageCode='en'
    )
    
    # Simple logic to redact based on results
    output = text
    for entity in response['Entities']:
        if entity['Score'] > 0.90: # Confidence threshold
            start = entity['BeginOffset']
            end = entity['EndOffset']
            pii_type = entity['Type']
            
            # Replace target slice with its type label
            output = output.replace(text[start:end], f"[{pii_type}]")
            
    return output

# Example Usage
raw = "Call me at 555-0199 or email john.doe@example.com."
print(clean := mask_pii(raw))
# Result: Call me at [PHONE] or email [EMAIL].

---

Knowledge Check: Test Your Privacy Knowledge

---

Summary

Security isn't a "vibe"; it's an architecture. By masking PII, using Guardrails, and ensuring data residency, you build a system that customers can trust. In the next lesson, we look at the specific Compliance Frameworks (GDPR, HIPAA) and the AWS tools that help you audit them.

---

Next Lesson: The Rulebook: Compliance Frameworks and AWS Tools