Identifying Risks to AI Solutions PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   Which of the following is an example of a prompt injection attack?

   • AA user submits a prompt in a different language to test the model's translation capabilities
   • BA user submits a very long prompt that causes the model to time out
   • CA user submits a prompt with misspelled words to see if the model corrects them
   • DA user includes text in a prompt that instructs the model to ignore its system instructions and reveal hidden system prompts
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A hospital's research team uses a Vertex AI PaLM API to analyze unstructured clinical notes. They have a strict requirement to ensure that no Protected Health Information (PHI) is sent to the model. The team currently relies on manual review of the notes before submission. They want to automate this with a Google Cloud service that can inspect the text and redact PHI before it is sent to the LLM. Which service and integration pattern is the most appropriate?

   • AConfigure a VPC Service Controls perimeter around the PaLM API and the Cloud Storage bucket containing the notes.
   • BUse Vertex AI Model Armor's data leakage prevention filters to redact PHI in the model's response.
   • CTrain a custom classification model on Vertex AI to identify PHI and use it to filter the notes before submission.
   • DUse Cloud DLP's `inspect` and `deidentify` methods in a Cloud Function that processes the text before calling the PaLM API.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A healthcare startup uses a Vertex AI-hosted model to process patient records. The security team is concerned about a compromised service account being used to extract the full training dataset via repeated prediction requests. Which combination of measures would most directly mitigate this specific data exfiltration risk?

   • ARotate the service account key every 24 hours and require all requests to include a short-lived OAuth token.
   • BApply a safety filter to the model's output to block any response containing patient identifiers.
   • CEnable VPC Service Controls and set per-user rate limits on the prediction endpoint, while logging all access patterns to Cloud Logging.
   • DEncrypt the training dataset with a customer-managed key and store the key in a different project.
   Reveal Solution Discussion Select an answer first
5. 19expert · hard

   A multinational corporation uses a Vertex AI model to process employee HR data. The data is stored in a BigQuery dataset in the `us-central1` region. The security team has identified a risk that a malicious insider with `bigquery.dataViewer` access could copy the training data to a personal project outside the organization. The company wants to enforce a technical control that prevents this specific cross-project data movement. Which control is the most targeted and effective?

   • AExport the BigQuery dataset to a Cloud Storage bucket in the same project and set the bucket to `uniform` bucket-level access.
   • BEnable Cloud DLP inspection on the BigQuery dataset and set up a job to redact all employee names and IDs.
   • CCreate a VPC Service Controls perimeter that includes the BigQuery dataset and the Vertex AI model, and deny all egress to projects outside the perimeter.
   • DRevoke the `bigquery.dataViewer` role from the insider and grant them a custom role with only `bigquery.jobs.create`.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   A machine learning engineer wants to detect unauthorized extraction of training data from a deployed model. Which monitoring approach directly addresses this risk?

   • AEncrypting the model artifacts at rest using customer-managed encryption keys
   • BSetting up alerts for unusual patterns of API access, such as high-volume or repetitive queries from a single source
   • CRestricting access to the model's source code repository to authorized personnel only
   • DRegularly retraining the model on updated data to improve accuracy
   Reveal Solution Discussion Select an answer first