AI Security Architecture and Controls CRAGE Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A security architect is designing an AI system for autonomous vehicles. The system must be highly available and safe. The architect is considering (1) a redundant model ensemble that votes on decisions, and (2) a single model with extensive adversarial training. The system has limited computational resources on the vehicle. Which approach better addresses availability and safety?

   • AA redundant model ensemble, because it provides failover if one model fails and can reduce the impact of adversarial inputs through voting.
   • BA single model with adversarial training, because it is simpler and uses fewer computational resources.
   • CNeither, because autonomous vehicles should not rely on AI for safety-critical decisions.
   • DBoth, because they address different threats and should be combined.
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   A global bank is deploying an AI system to detect money laundering. The system must comply with GDPR for EU customer data and with local data-residency laws that require EU customer data to remain within the EU. The bank also needs to collaborate with a non-EU subsidiary to improve the model. The data science team proposes using federated learning with differential privacy. Which consideration is most critical for the bank to address?

   • AEnsure that model updates shared with the non-EU subsidiary do not contain personal data and that any transfer complies with GDPR
   • BStore all model updates in the EU and only share aggregated statistics
   • CObtain explicit consent from every customer for international data transfer
   • DUse homomorphic encryption to encrypt all model updates before sharing
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A healthcare startup is deploying a diagnostic image classification model. During red-team testing, a security analyst demonstrates that adding a subtle, imperceptible pattern to an X-ray image causes the model to misclassify a malignant tumor as benign. The startup needs to reduce this risk before production. Which control should they implement first?

   • AAdd adversarial training examples to the training dataset and retrain the model.
   • BDeploy the model behind a web application firewall that blocks unusual image uploads.
   • CRequire two-person review for all model predictions above a confidence threshold.
   • DEncrypt the model weights at rest using a hardware security module.
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A security architect is designing an AI system for a hospital that predicts patient deterioration. The system must be highly available because it is used in real-time clinical settings. The architect is considering two designs: (A) a single on-premises server with a failover backup, and (B) a cloud-based deployment with auto-scaling and multi-region redundancy. Which factor is most important in deciding between the two?

   • AThe cloud-based deployment offers higher availability due to redundancy, but the hospital must ensure network connectivity and data-residency compliance
   • BThe on-premises server is always more secure than the cloud
   • CThe on-premises server provides better explainability than the cloud
   • DThe cloud deployment is always cheaper than on-premises
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which mechanism directly supports the transparency of an AI system by providing human-understandable reasons for its decisions?

   • AExplainability
   • BModel documentation
   • CAuditability
   • DDifferential privacy
   Reveal Solution Discussion Select an answer first