A financial enterprise deploys custom-trained classification models to Vertex AI to predict sensitive loan outcomes. A cloud security engineer is tasked with establishing security controls across the model deployment pipeline and inference serving layer to meet the following requirements: Model Governance & Integrity: Maintain centralized lifecycle management with distinct versioning in Vertex AI Model Registry, and enforce cryptographic verification of model artifacts prior to deployment to prevent tampering. Privacy & Attack Mitigation: Prevent model inversion and membership inference attacks that attempt to reconstruct private training data from the prediction API responses. Network Isolation: Ensure client workloads within the enterprise Virtual Private Cloud (VPC) reach the inference endpoint privately without routing over the public internet. Which combination of controls should the security engineer implement?

Register versions in Vertex AI Model Registry; verify cryptographic hashes or digital signatures of the model artifact before deployment; implement Custom Prediction Routines (CPR) to sanitize prediction outputs and restrict full probability distributions; and deploy the model to Vertex AI private endpoints connected to the VPC.: ### What the Solution Is This architecture establishes end-to-end model governance, cryptographic provenance verification, API output defenses, and private network perimeter security using native **Vertex AI** capabilities. ### How It Meets the Requirements - **Model Governance & Integrity:** **Vertex AI Model Registry** acts as the central, searchable repository to track, version, and manage model lifecycles. Cryptographically hashing or signing model artifacts ensures that only tamper-proof, verified models that match the approved training lineage are deployed into production endpoints. - **Privacy & Attack Mitigation (Model Inversion & Membership Inference):** Model inversion and membership inference attacks exploit overly detailed prediction outputs (such as high-precision confidence score vectors across all classes) to infer training records or reconstruct sensitive features. By deploying **Custom Prediction Routines (CPR)**, engineers can implement custom pre/post-processing logic directly within the model container to filter or round probabilities, return only top-1 predicted labels, or apply perturbation/differential privacy mechanisms. - **Network Isolation:** Deploying the model to **Vertex AI private endpoints** establishes direct peer-to-peer connectivity with the enterprise **Virtual Private Cloud (VPC)**, eliminating exposure to internet routing. ### Key Advantages - **Centralized Model Lifecycle:** Model Registry simplifies tracking multiple versions and rollback operations. - **Tamper Resistance:** Pre-deployment cryptographic verification validates artifact provenance and integrity. - **Output Hardening:** CPR post-processing removes sensitive metadata and excessive probability precision without modifying the underlying model weights. - **Zero Public Exposure:** Private endpoints keep all inference traffic strictly inside the local Google Cloud network. ### Why This Is Optimal This solution directly addresses data privacy vulnerabilities at the inference tier while enforcing supply-chain integrity controls and private network transport.

Vertex Ai Model Registry Integrity Privacy Controls | Professional Cloud Security Engineer

Vertex Ai Model Registry Integrity Privacy Controls

Worked example. The correct answer is already marked and every option is explained below, so there is nothing to select here. To answer questions yourself, start the free trial.

A financial enterprise deploys custom-trained classification models to Vertex AI to predict sensitive loan outcomes. A cloud security engineer is tasked with establishing security controls across the model deployment pipeline and inference serving layer to meet the following requirements:

Model Governance & Integrity: Maintain centralized lifecycle management with distinct versioning in Vertex AI Model Registry, and enforce cryptographic verification of model artifacts prior to deployment to prevent tampering.
Privacy & Attack Mitigation: Prevent model inversion and membership inference attacks that attempt to reconstruct private training data from the prediction API responses.
Network Isolation: Ensure client workloads within the enterprise Virtual Private Cloud (VPC) reach the inference endpoint privately without routing over the public internet.

Which combination of controls should the security engineer implement?

Register versions in Vertex AI Model Registry; verify cryptographic hashes or digital signatures of the model artifact before deployment; implement Custom Prediction Routines (CPR) to sanitize prediction outputs and restrict full probability distributions; and deploy the model to Vertex AI private endpoints connected to the VPC.

Register models using BigQuery ML remote model references in Vertex AI Model Registry; apply Model Armor response templates to filter raw classification probabilities; and configure standard public endpoints protected exclusively by IAM authentication.

Import models into Vertex AI Model Registry without custom routines; configure Dataflow workers to evaluate ground truth batch predictions continuously; and store model binaries in multi-region Cloud Storage buckets with Object Versioning.

Deploy models to Google Distributed Cloud using the google_ml_integration extension; store credentials in Secret Manager using sm_secret; and rely on database-level transform functions to truncate inference outputs over standard external connections.

Professional Cloud Security Engineer

Vertex Ai Model Registry Integrity Privacy Controls

Related questions