Secure Generative AI: Bedrock Guardrails and Tokenization

This article details integrating Amazon Bedrock Guardrails with third-party tokenization services, like Thales CipherTrust, for enhanced data security in generative AI applications. Bedrock Guardrails detect and mask sensitive data (PII) in prompts and responses, but this masking can hinder data reversibility. Tokenization offers a solution by replacing sensitive data with format-preserving tokens, mathematically unrelated to the originals but easily reversible by authorized systems. The integration leverages Bedrock Guardrails’ ApplyGuardrail API, separating content assessment from model invocation. This allows for tokenization processing before model invocation and after the response, ensuring data protection without sacrificing functionality. A financial advisory application example illustrates the workflow: a customer gateway service sends a query; the ApplyGuardrail API detects PII; a tokenization service replaces PII with tokens; the financial analysis engine processes the tokenized data; and a response processing service detokenizes the results before presenting them to the customer. AWS Step Functions orchestrates this process using Lambda functions. The solution requires an AWS account with Bedrock and IAM permissions, a Python environment with boto3, and a deployed tokenization service. The article provides code examples for creating guardrails, applying them, invoking the tokenization service (Thales CipherTrust in this case), and replacing masks with tokens. The process ensures data confidentiality while maintaining data utility for authorized systems, offering a robust solution for organizations in highly regulated industries.

(Source: https://aws.amazon.com/blogs/machine-learning/integrate-tokenization-with-amazon-bedrock-guardrails-for-secure-data-handling/)