3.3 Securing AI Workloads PROFESSIONAL-CLOUD-SECURITY-ENGINEER Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   An attacker repeatedly queries a deployed ML model and uses the responses to create a functional replica of the model. Which type of attack is this?

   • AMembership inference
   • BData poisoning
   • CModel inversion
   • DModel stealing
   Reveal Solution Discussion Select an answer first
2. 12application · hard

   A hospital deploys a Gemini Enterprise agent to help doctors query patient records. The compliance team requires that the agent's responses must not contain protected health information (PHI) unless the doctor explicitly requests it. What is the most effective way to enforce this?

   • ARestrict access to the agent using IAM and require that only doctors with a specific role can query patient data.
   • BEnable Cloud DLP inspection on the agent's responses and redact any detected PHI before it is shown to the doctor.
   • CConfigure the agent's grounding with a data store that contains only de-identified summaries of patient records.
   • DSet the agent's temperature to 0.0 to ensure deterministic and conservative responses.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A tech company has deployed a large language model (LLM) for internal code generation. The security team wants to detect if the model is being used to generate malicious code or if it is leaking proprietary source code in its responses. What should they set up?

   • ARestrict the model's access to the internet and only allow it to be accessed from the internal corporate network.
   • BUse Cloud IDS to monitor network traffic to and from the model endpoint for signs of data exfiltration.
   • CEnable Cloud Logging for the model endpoint and export the logs to BigQuery for analysis with a custom query that flags suspicious keywords.
   • DDeploy a Web Application Firewall (WAF) in front of the model endpoint to filter out requests containing known malicious payloads.
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A security team is evaluating the risk of model stealing for a proprietary image classification model deployed on Vertex AI. They have identified that an attacker could use the model's prediction API to train a substitute model. The team must choose a mitigation that balances security with the need to maintain a high-quality API for legitimate customers. What is the most balanced approach?

   • AImplement rate limiting and monitor for unusual query patterns, such as a high volume of requests with similar inputs.
   • BAdd differential privacy noise to the model's confidence scores before returning them to the client.
   • CRestrict access to the API by requiring an API key and only issuing keys to vetted customers.
   • DWatermark the model's outputs by embedding a subtle, unique pattern in the predictions.
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A financial services firm deploys a fraud-detection model. The security team is concerned about adversarial attacks where subtle perturbations to transaction features cause the model to misclassify fraudulent activity as legitimate. They want to detect such attacks in production. What should they implement?

   • AEnable Cloud Logging for the model endpoint and set up alerts for unusual query patterns, such as high request rates from a single IP.
   • BStore all inference requests in Cloud Storage with a retention policy of 90 days for post-incident forensic analysis.
   • CDeploy a separate adversarial-detection model that scores each input for perturbation likelihood before the primary model processes it.
   • DRestrict access to the model's confidence scores and only return the final classification label to the application.
   Reveal Solution Discussion Select an answer first