AI Security Architecture and Controls CRAGE Practice Questions (Page 6)

Questions 26–30

1. 26foundation · easy

   How does integrating AI security architecture with governance frameworks primarily benefit an organization?

   • AIt automatically prevents all adversarial attacks
   • BIt ensures that security controls align with regulatory requirements and organizational policies
   • CIt guarantees that AI models are 100% accurate
   • DIt eliminates the need for separate data privacy measures
   Reveal Solution Discussion Select an answer first
2. 27application · medium

   A bank's AI credit-scoring model must provide explanations to customers who are denied credit. The model is a complex gradient-boosting model. The compliance team also requires that the explanations be consistent and reproducible for audits. Which approach should the bank take?

   • AUse SHAP (SHapley Additive exPlanations) to generate feature-attribution explanations for each decision and log them with the input data.
   • BUse a model-agnostic explainer like LIME but do not store the explanations to save storage.
   • CProvide customers with the model's overall feature importance scores instead of individual explanations.
   • DReplace the gradient-boosting model with a simple linear regression model to make explanations easier.
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A security team is threat-modeling an AI-based autonomous vehicle system. They identify two major threats: (1) an attacker can physically place stickers on road signs to cause misclassification, and (2) an attacker can tamper with the training data to introduce a backdoor. The team has limited budget and must choose a single control that addresses both threats. Which control is most effective?

   • ARestrict access to the training data and use anomaly detection on the training pipeline
   • BDeploy an ensemble of multiple models and average their predictions
   • CImplement robust adversarial training that includes both physical-world perturbations and backdoor triggers
   • DUse input sanitization to detect and remove stickers from camera images
   Reveal Solution Discussion Select an answer first
5. 29foundation · easy

   Which security control is most important during the deployment stage of an AI model to protect against inference-time attacks?

   • AApplying differential privacy to the training data
   • BRegularly retraining the model with new data
   • CEncrypting the training dataset at rest
   • DRobust input validation and anomaly detection on incoming requests
   Reveal Solution Discussion Select an answer first
6. 30foundation · easy

   Which principle of AI security architecture is primarily concerned with ensuring that the outputs of an AI model have not been silently altered or corrupted?

   • AAvailability
   • BIntegrity
   • CConfidentiality
   • DNon-repudiation
   Reveal Solution Discussion Select an answer first