AI Security Architecture and Controls CRAGE Practice Questions (Page 9)

Questions 41–45

1. 41application · medium

   An e-commerce company uses a recommendation model that is critical for revenue. The security team is designing an architecture to ensure the model remains available during a DDoS attack on the public API. Which control is most directly aligned with the availability principle?

   • ADeploy the model behind a load balancer with auto-scaling and rate limiting
   • BUse a private endpoint for all model access
   • CEncrypt all model inputs and outputs in transit
   • DStore model snapshots in an immutable object store
   Reveal Solution Discussion Select an answer first
2. 42application · medium

   A city government wants to release an AI model trained on citizen data for public research. The legal team requires that the model not reveal information about any individual citizen. Which technique should be applied to the model or its training process?

   • ARemove all personally identifiable information from the training data before training.
   • BTrain the model with differential privacy, adding calibrated noise to the training process.
   • CUse homomorphic encryption to encrypt the model before releasing it.
   • DRelease only the model's accuracy metrics instead of the model itself.
   Reveal Solution Discussion Select an answer first
4. 43expert · hard

   A company is developing an AI system for medical diagnosis. The system must be explainable to doctors and auditable by regulators. The team is considering using a complex deep learning model with post-hoc explanations versus a simpler interpretable model. Which approach is more likely to satisfy both explainability and auditability requirements?

   • AUse a deep learning model with SHAP explanations, because it is more accurate
   • BUse a deep learning model and provide only the model's confidence score
   • CUse a simpler interpretable model, because its decisions are inherently explainable and easier to audit
   • DUse a deep learning model and rely on external auditors to explain it
   Reveal Solution Discussion Select an answer first
5. 44application · medium

   A research institute wants to publish a statistical analysis of a sensitive dataset while ensuring that individual records cannot be re-identified. The dataset contains many unique combinations of attributes. Which technique should they apply?

   • ASimple anonymization by removing names and social security numbers
   • BDifferential privacy with a calibrated privacy budget
   • CEncrypting the dataset and publishing the encryption key
   • DUsing federated learning to train a model on the data
   Reveal Solution Discussion Select an answer first
6. 45application · medium

   A security team is threat modeling an AI system that uses a public pre-trained language model. They are concerned that an attacker could craft inputs that cause the model to output sensitive information from its training data. Which attack is this, and which control is most relevant?

   • APrompt injection attack; control by implementing input validation.
   • BModel inversion attack; control by encrypting the model weights.
   • CTraining data extraction attack; control by filtering and sanitizing the model's outputs.
   • DData poisoning attack; control by monitoring the training data pipeline.
   Reveal Solution Discussion Select an answer first