AWS re:Invent 2025 - Data protection strategies for AI data foundation (AIM339)

AWS re:Invent 2025 - Data Protection Strategies for AI Data Foundation

Overview

This presentation from AWS re:Invent 2025 focuses on strategies for securing and protecting sensitive data used in AI and machine learning applications, particularly in the context of a nonprofit healthcare chatbot. The speakers, Derek Martinez and Sabrina Petruso, outline a comprehensive data protection framework and demonstrate the implementation of key security and privacy controls through a live coding example.

Key Challenges Addressed

• Handling sensitive patient data in AI/ML applications
• Mitigating risks of prompt injection attacks on language models
• Ensuring data privacy and compliance (e.g. HIPAA) in AI data pipelines

Data Protection Framework

The presenters outline a 6-layer data protection framework:

1. Encryption

• Encrypt data at rest and in transit using AWS KMS

2. Fine-Grained Access Control

• Leverage IAM to implement least-privilege access controls

3. Auditing and Monitoring

• Use AWS CloudTrail to log and audit all actions

4. Automated Compliance

• Leverage AWS Config to define and monitor compliance rules (e.g. HIPAA)

5. PII Detection and Sanitization

• Utilize Amazon Textract and Amazon Comprehend to detect and mask PII
• Implement differential privacy techniques like k-anonymity and randomization

6. Prompt Injection Defense

• Detect and mitigate potential prompt injection attacks on the chatbot

Live Coding Example

The presenters walk through a live coding example of the data protection pipeline built on Amazon SageMaker:

1. Data Ingestion

• Internal data owner uploads raw patient data to an S3 bucket

2. Data Processing

• Amazon Textract extracts text from documents
• Amazon Comprehend detects and masks PII using differential privacy
• Processed data is stored in a separate S3 bucket

3. Prompt Injection Defense

• API Gateway exposes a backend Lambda function to process user prompts
• Lambda function checks for and mitigates potential prompt injection attacks
• AWS Config applies HIPAA-specific compliance rules

4. Auditing and Monitoring

• All actions are logged to a secure S3 bucket via AWS CloudTrail
• Audit logs track PII detection, masking, and prompt injection events

Business Impact and Use Cases

• Enables secure and compliant use of sensitive data in AI/ML applications
• Protects against data breaches and privacy violations
• Allows organizations to leverage AI while maintaining data governance
• Particularly relevant for industries like healthcare, finance, and government

Key Takeaways

• Comprehensive data protection is crucial for responsible AI development
• Layered security approach with encryption, access control, auditing, and compliance is essential
• Differential privacy techniques like k-anonymity can effectively mask PII
• Prompt injection is a significant risk for language model-based applications
• The presented framework and code can be adapted for various AI use cases