Lesson 11: AI Security, Data Privacy & Compliance on AWS

1. AI Security on AWS — Overview

AI workloads require security at multiple layers: data, model, infrastructure, and application.

AI Security Layers:
┌─────────────────────────────────────────┐
│  APPLICATION SECURITY                   │
│  Guardrails, input validation,          │
│  prompt injection prevention            │
├─────────────────────────────────────────┤
│  MODEL SECURITY                         │
│  Model access control, versioning,      │
│  model integrity, adversarial defense   │
├─────────────────────────────────────────┤
│  DATA SECURITY                          │
│  Encryption, PII handling, data access  │
│  control, audit logging                 │
├─────────────────────────────────────────┤
│  INFRASTRUCTURE SECURITY                │
│  VPC, security groups, endpoints,       │
│  network isolation, IAM                 │
└─────────────────────────────────────────┘

2. IAM for AI/ML Workloads

2.1. IAM for Amazon Bedrock

Permission	What it controls
bedrock:InvokeModel	Call a specific FM
bedrock:InvokeModelWithResponseStream	Streaming model invocation
bedrock:CreateKnowledgeBase	Create RAG knowledge bases
bedrock:CreateGuardrail	Create safety guardrails
bedrock:CreateModelCustomizationJob	Start fine-tuning jobs

2.2. IAM Best Practices for AI

• Least privilege: Grant only required permissions per role
• Separate roles: Different roles for data scientists, ML engineers, admins
• Resource-based policies: Restrict access to specific models
• Service-linked roles: Let AWS services assume roles as needed
• Condition keys: Restrict by IP, VPC, time, MFA

Exam tip: "How to restrict which foundation models a team can use in Bedrock?" → IAM policy with bedrock:InvokeModel and resource ARN for specific models.

3. Data Encryption

3.1. Encryption at Rest

Service	Default Encryption	Custom Key (KMS)
Amazon Bedrock	AWS-managed key	Customer-managed KMS key
SageMaker training	AWS-managed key	Customer-managed KMS key
SageMaker notebooks	AWS-managed key	Customer-managed KMS key
S3 (training data)	SSE-S3	SSE-KMS, SSE-C
Knowledge Bases vectors	Depends on vector DB	KMS encryption supported

3.2. Encryption in Transit

• All AWS API calls use TLS 1.2+ by default
• Bedrock API calls are encrypted in transit
• SageMaker endpoints use HTTPS
• Inter-node training communication: encrypted

3.3. AWS KMS for AI

Use AWS KMS when you need:
✓ Control over encryption keys
✓ Key rotation policies
✓ CloudTrail audit of key usage
✓ Cross-account key sharing
✓ Compliance requirements (HIPAA, PCI-DSS)

4. Network Security

4.1. VPC Endpoints for AI Services

VPC endpoints allow private access to AWS AI services — traffic does not traverse the internet.

Without VPC Endpoint:
App in VPC → Internet Gateway → Public Internet → Bedrock API

With VPC Endpoint (PrivateLink):
App in VPC → VPC Endpoint → AWS Private Network → Bedrock API
             (no internet!)

Service	VPC Endpoint Type
Amazon Bedrock	Interface (PrivateLink)
SageMaker Runtime	Interface (PrivateLink)
SageMaker API	Interface (PrivateLink)
Amazon S3	Gateway or Interface

4.2. SageMaker Network Isolation

• VPC mode: Run training/inference inside your VPC
• Network isolation: No internet access for containers (EnableNetworkIsolation=true)
• Security groups: Control inbound/outbound traffic
• Private subnets: No direct internet access

Exam tip: "How to ensure Bedrock API calls don't traverse the public internet?" → VPC endpoint (AWS PrivateLink) for Amazon Bedrock.

5. PII Detection & Data Privacy

5.1. PII Detection Services

Service	PII Capability	Data Type
Amazon Comprehend	Detect and redact PII entities	Text
Amazon Macie	Discover PII in S3 buckets	Files in S3
Bedrock Guardrails	Block/anonymize PII in FM I/O	FM prompts/responses
AWS Glue DataBrew	PII detection in data pipelines	Structured data

5.2. Common PII Types for Exam

PII Type	Examples
Direct identifiers	Name, SSN, email, phone, passport number
Financial	Credit card number, bank account, tax ID
Health	Medical record number, health conditions (PHI)
Location	Home address, GPS coordinates
Digital	IP address, device ID, login credentials

6. Amazon Bedrock Security

6.1. Data Privacy in Bedrock

• Data isolation: Your data is NOT used to train base FMs
• Data stays in region: Processed in the AWS region you choose
• Custom models: Fine-tuned models are private to your account
• No data sharing: Your prompts/responses are not shared with model providers
• Encryption: All data encrypted at rest and in transit

6.2. Monitoring & Logging

Service	What it logs
AWS CloudTrail	API calls (who invoked which model, when)
Amazon CloudWatch	Model invocation metrics (latency, errors, tokens)
Bedrock Model Invocation Logging	Full prompts and responses (to S3 or CloudWatch)

Exam tip: "How to audit which users are calling Bedrock models?" → CloudTrail. "How to log the actual prompts and responses?" → Bedrock Model Invocation Logging.

7. Compliance & Governance

7.1. AWS Compliance Programs

Program	What	Relevant AI Services
SOC 1/2/3	Security controls audit	Bedrock, SageMaker
HIPAA	Healthcare data protection	SageMaker, Comprehend Medical
GDPR	EU data privacy	All AWS services (data residency)
PCI-DSS	Payment card data security	SageMaker (with controls)
FedRAMP	US government cloud security	GovCloud regions
ISO 27001	Information security management	Bedrock, SageMaker

7.2. Shared Responsibility Model for AI

CUSTOMER responsibility ("Security IN the cloud"):
├── Training data quality and bias
├── Model selection and evaluation
├── Prompt design and guardrails configuration
├── IAM permissions and access control
├── PII handling and data classification
├── Application-level security
└── Compliance with industry regulations

AWS responsibility ("Security OF the cloud"):
├── Physical infrastructure security
├── Network and hardware security
├── Base FM provider management
├── Service availability and reliability
├── Encryption implementation
└── Compliance certifications

8. Data Governance for AI

Practice	AWS Service
Data cataloging	AWS Glue Data Catalog
Data classification	Amazon Macie
Access control	AWS Lake Formation
Data lineage	SageMaker ML Lineage Tracking
Data quality	SageMaker Data Wrangler, Glue DataBrew

9. Practice Questions

Q1: A financial services company wants to use Amazon Bedrock but requires that API calls do NOT traverse the public internet. What should they configure?

• A) AWS Direct Connect
• B) VPC endpoint (AWS PrivateLink) for Bedrock ✓
• C) VPN connection
• D) CloudFront distribution

Explanation: A VPC interface endpoint (PrivateLink) for Amazon Bedrock allows private connectivity from within a VPC without traffic going through the internet.

Q2: According to the AWS Shared Responsibility Model, who is responsible for ensuring training data does not contain bias?

• A) AWS
• B) The foundation model provider
• C) The customer ✓
• D) Both AWS and the customer equally

Explanation: Under the shared responsibility model, customers are responsible for "security IN the cloud" — this includes training data quality, bias detection, and ethical AI practices. AWS is responsible for infrastructure security.

Q3: A company needs to discover which S3 buckets contain personally identifiable information (PII) before using the data for ML training. Which service should they use?

• A) Amazon Comprehend
• B) Amazon Macie ✓
• C) Amazon Inspector
• D) AWS Config

Explanation: Amazon Macie uses machine learning to automatically discover and classify sensitive data (including PII) stored in Amazon S3 buckets. Comprehend detects PII in text, but Macie is designed for S3-level discovery.