# RAISP02-BP02 Privacy: Build privacy-preserving mechanisms into the core AI system
<a name="raisp02-bp02"></a>

 Design your system from the start to protect confidential and personal data. This may include incorporating techniques like data encryption, access controls, data minimization, data obfuscation, and privacy-preserving training methods directly into how your system works, based on your release criteria. 

 For example, if your release criteria include keeping certain types of user information confidential, you might build in automatic data masking, use techniques that scramble sensitive information while keeping it useful for training, or set up your system to process information without storing sensitive details. The specific privacy mechanisms you choose should align with your release criteria. 

 **Level of risk exposed if this best practice is not established:** High 

## Implementation considerations
<a name="implementation-considerations-65"></a>

1.  Implement data protection measures: Establish security protections around sensitive information. First, identify essential data requirements through data minimization analysis. Create a mapping of sensitive fields and implement anonymization. For example, in a healthcare system, converting 'John Doe, diabetic, 123 Main Street' to 'Patient\_2384, condition\_type\_2, region\_14' maintains analytical value while protecting individual privacy. Encrypt sensitive data at rest and in transit. Establish role-based access controls with documented access levels for sensitive data. 

1.  Apply privacy-preserving training techniques: Consider using differential privacy techniques to introduce controlled noise to the training process. For example, when processing customer transaction data, apply calculated variations to individual purchases while maintaining accurate aggregate patterns. Consider using federated learning to enable distributed model training where data remains at source locations. 

 For example, with federated learning, healthcare institutions can improve diagnostic models by sharing only parameter updates instead of raw patient data. Consider using gradient clipping to block individual training examples from disproportionately influencing model learning, maintaining both privacy and model quality. 

## Resources
<a name="resources-62"></a>

 **Related documents:** 
+  [Differentially Private Fair Learning](https://arxiv.org/abs/1812.02696) 
+  [Remove PII from conversations by using sensitive information filters](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-sensitive-filters.html) 
+  [ISO/IEC 42001:2023 A.6.1.2 Objectives for responsible development of AI system](https://www.iso.org/standard/42001) 

 **Related video:** 
+  [Amazon Bedrock Guardrails: Implementing Custom Safeguards for Responsible AI Applications](https://aws.amazon.com/awstv/watch/02103dd95d3/) 
+  [AWS re:Inforce 2025 - Privacy-first generative AI: Establishing guardrails for compliance (COM224)](https://www.youtube.com/watch?v=GAjWNoxgkYY) 

 **Related tools:** 
+  [Amazon Bedrock Guardrails and PII removal](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-sensitive-filters.html)