GCP Professional ML Engineer Practice Question 12: PII Deidentification Before Training
Google Cloud Professional Machine Learning Engineer · Test 2 · question 2 of 10
Support-ticket text is being prepared as training data for a ticket-routing classifier. The raw tickets contain customer names, email addresses and payment card numbers typed into the free-text body. Governance requires that no direct identifier reaches the training bucket, while the surrounding wording must stay intact because it carries the routing signal.
What should you do before the data is written to the training bucket?
Show answer and explanation
Answer:
Sensitive Data Protection (the service that includes the Cloud Data Loss Prevention API) inspects content for built-in infoTypes such as PERSON_NAME, EMAIL_ADDRESS and CREDIT_CARD_NUMBER, then applies a de-identification transformation (masking, redaction, replacement with the infoType name, or cryptographic tokenization) to just the matched spans. The identifiers never reach the training bucket while the rest of the ticket wording is preserved, which is what the routing signal depends on (option B).
- A is wrong because access control decides *who* may read the identifiers. It does not remove them, so the identifiers still land in the training data and in the model's training corpus. - C is wrong because encryption at rest is transparent to anyone authorized to read the object. The plaintext identifiers are still in the training set the moment the pipeline reads it. - D is wrong because discarding whole tickets throws away exactly the records that contain the most free-text detail, shrinking and biasing the training set instead of sanitizing it.
Reference: De-identifying sensitive data