Keeping PII out of support transcripts used for fine-tuning
MLA-C02
Data Preparation for ML and AI
Medium
Question 22 of 30
A bank wants to fine-tune a foundation model on 2 million customer support chat transcripts that are stored as text files in Amazon S3. Customer names, street addresses, and card numbers appear anywhere in the free-form sentences. The data scientists who prepare the training dataset must not see these values. The bank wants a managed approach that needs minimal custom code.
Which solution will meet these requirements?
Show answer and explanation
BRun an Amazon Comprehend asynchronous PII redaction job that writes redacted copies to a separate prefix
An Amazon Comprehend PII redaction job meets the requirements. Comprehend detects PII entities such as NAME, ADDRESS and CREDIT_DEBIT_NUMBER in English or Spanish text, and an asynchronous job returns a copy of each document with every detected entity redacted, so the data scientists work only from the redacted prefix. Amazon Macie discovers and reports sensitive data in S3 by generating findings, but it does not produce redacted copies, so the data scientists would still read the originals. Regular expressions for card number patterns miss names and addresses written in free-form sentences. KMS encryption protects data at rest, but anyone granted decrypt access reads the plaintext, PII included.
Source: docs.aws.amazon.com/
Anonymous, no account needed. Your tally stays in this browser.