Question

Your team deployed a Vertex AI model that predicts patient readmission risk using clinical records containing protected health information (PHI). During a compliance review, auditors ask you to demonstrate that access to model predictions and the underlying request payloads is controlled, traceable, and that PHI is not inadvertently exposed to unauthorized team members through monitoring artifacts. Which approach best addresses this responsible-AI privacy risk?

Accepted Answer

Restrict endpoint and logging-dataset access with IAM least-privilege roles, enable Cloud Audit Logs for data-access events, and de-identify or exclude PHI fields before they are written to prediction logs.

Answer

Enable Vertex AI request-response logging to BigQuery with full payloads, then grant all data scientists BigQuery Viewer so they can debug predictions quickly.

Answer

Disable all prediction logging entirely so no PHI is ever stored, and rely on the model's aggregate accuracy metrics for any future audits.

Answer

Encrypt the model artifact with a customer-managed encryption key (CMEK) and consider the privacy requirement fully satisfied since the data is encrypted at rest.