GCP Professional Cloud Security Engineer Practice Question

FinServe Inc. stores card-holder data in BigQuery and trains a fraud-detection model with Vertex AI. The security team must (1) stop engineers or code running in Vertex AI from copying training data to resources outside the company's network, and (2) minimize the chance that online prediction requests could be used to reconstruct individual card numbers. Which two controls, used together, best satisfy both requirements?

Require engineers to access BigQuery only through the Cloud SQL Auth proxy and encrypt the training data with customer-managed encryption keys (CMEK) in Cloud KMS.
Create a VPC Service Controls perimeter that includes all BigQuery datasets and Vertex AI resources, and de-identify the card data with Cloud Sensitive Data Protection using format-preserving tokenization before training.
Disable external IP addresses on Vertex AI Workbench notebooks and apply BigQuery row-level access policies to restrict engineers to specific rows.
Encrypt the Vertex AI model artifacts with customer-supplied encryption keys (CSEK) and rotate the keys every 90 days.

Report Issue

Answer Description

Placing all BigQuery and Vertex AI resources inside the same VPC Service Controls service perimeter prevents both users and workloads from moving data to services or projects outside the perimeter, closing off common exfiltration paths. De-identifying the PCI data with Cloud Sensitive Data Protection (DLP) using format-preserving encryption or tokenization removes the real card numbers from the training set, so even successful model-inversion or membership-inference attacks cannot disclose them. Using CMEK or disabling external IPs are good hardening steps, but they do not themselves block data exfiltration or remove sensitive values from the model. Row-level security and notebook hardening help with least-privilege access, yet they do not mitigate inference attacks that exploit trained models to recover original data. Therefore, combining a VPC Service Controls perimeter with DLP-based tokenization is the most effective way to meet both security objectives.

Ask Bash

Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.