AI Security Architecture and Controls CRAGE Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A team is threat modeling an AI-based recommendation system. They identify that an attacker could submit crafted user profiles to influence the model's recommendations for other users. Which type of attack is this, and which lifecycle stage should be the primary focus for mitigation?

   • AModel inversion attack; mitigation should focus on the model storage layer.
   • BPrompt injection attack; mitigation should focus on the model's input validation.
   • CData poisoning attack; mitigation should focus on the training data pipeline.
   • DAdversarial example attack; mitigation should focus on the inference API.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A healthcare consortium wants to train a model on patient data from multiple hospitals. The hospitals have different data formats and privacy policies. The consortium is considering federated learning with differential privacy. Which trade-off must they carefully manage?

   • ADifferential privacy eliminates the need for data-sharing agreements
   • BFederated learning requires all hospitals to use the same cloud provider
   • CHomomorphic encryption is required for federated learning to be secure
   • DThe privacy budget consumed by differential privacy may reduce model accuracy, and the consortium must balance privacy guarantees with model utility
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A team is building a spam filter. An attacker wants to make the filter classify all emails as 'not spam' by injecting a large number of legitimate-looking emails into the training data. Which type of attack is this, and which lifecycle stage should be protected?

   • AModel inversion during the deployment stage
   • BData poisoning during the training stage
   • CAdversarial perturbation during the inference stage
   • DPrompt injection during the monitoring stage
   Reveal Solution Discussion Select an answer first
5. 19expert · hard

   A company is deploying a large language model (LLM) chatbot for customer service. The security team is concerned about prompt injection attacks that could cause the chatbot to reveal sensitive information. They are considering two controls: (1) filtering user inputs for malicious patterns, and (2) restricting the chatbot's access to sensitive data. Which control is more effective as a primary defense?

   • AEncrypting the chatbot's responses, because it prevents data leakage
   • BFiltering user inputs for malicious patterns, because it prevents the attack from occurring
   • CRestricting the chatbot's access to sensitive data, because it limits the impact even if a prompt injection succeeds
   • DUsing both controls equally, because neither is sufficient alone
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A regulator requires that an AI system used for loan approvals provide explanations that are understandable to the applicant. The data science team has developed a highly accurate gradient boosting model. The team is debating between using SHAP explanations or switching to a simpler logistic regression model. Which consideration is most important?

   • ALogistic regression is always less accurate than gradient boosting
   • BSHAP is computationally expensive and cannot be used in production
   • CSHAP explanations may be approximate and not fully faithful to the model, while logistic regression provides inherently faithful explanations
   • DThe regulator will accept any explanation as long as it is generated by an algorithm
   Reveal Solution Discussion Select an answer first