Sensitive Information Filtering in Amazon Bedrock

Question

I'm trying to determine the best way to filter sensitive information when using  Amazon Bedrock Guardrails.  I really like the intuitive Guardrails configuration wizard; however, I notice there is no way to Mask user input that contains sensitive information - Masking is only available on model responses. To me, it would seem very helpful if we also had the capability to mask the input before it even gets sent to the model, but I'm still in the learning phase here.

Therefore, since I can't redact my input using out-of-box Guardrail functionality (I can only Block it), are my best options to perform my own custom Regex manipulation on the input before submitting to Bedrock, or maybe redact it via Amazon Comprehend first?

An example use case: I want to perform summarization on a conversation that may contain sensitive information, but I would like to redact any sensitive elements before it gets sent to any language model.

Thank you.

Michael Shapira · Accepted Answer

You can use Amazon Conprehend, but isn't sending the raw data to Comprehend the same as sending it to Amazon Bedrock and asking to return the same text without PII data?
You can implement any of the above as a preprocessing step before asking for summarization. 
Regular expression probably is less effective unless you know exactly what information you want to extract.

Sensitive Information Filtering in Amazon Bedrock

Add your answer

Relevant content