AI Threat and Vulnerability Management AAISM Practice Questions (Page 5)

Questions 21–25

1. 21foundation · easy

   What is the goal of a model extraction attack?

   • ATo create a copy of the model by querying it and observing its outputs
   • BTo cause the model to reveal its training data through its outputs
   • CTo inject malicious code into the model's training data
   • DTo overload the model with requests to cause a denial of service
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   Which attack vector involves inserting a hidden trigger into a model so that it behaves maliciously only when the trigger is present?

   • AData exfiltration
   • BBackdoor insertion
   • CEvasion attack
   • DModel inversion
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A financial services firm deploys a fraud-detection model that approves or declines transactions in real time. The security team notices a sudden increase in false declines for a specific merchant category, but the model's overall accuracy remains stable. Which action is most appropriate to investigate whether an adversarial attack is occurring?

   • AReview the distribution of input features for transactions in that merchant category for signs of adversarial perturbation.
   • BRetrain the model with more recent transaction data.
   • CDisable the model and switch to manual review for that merchant category.
   • DIncrease the model's decision threshold to reduce false declines.
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A security team is assessing an AI system that uses a pre-trained model from a third-party vendor. The team has limited resources and must decide where to focus its vulnerability assessment. Which approach is most effective?

   • AConduct a penetration test on the vendor's infrastructure.
   • BFocus on the data pipeline and the interface between the pre-trained model and the application, as these are the areas the organization controls.
   • CPerform a full source code review of the pre-trained model.
   • DAssume the pre-trained model is secure and only test the application layer.
   Reveal Solution Discussion Select an answer first
6. 25expert · hard

   A bank's credit-scoring model is suspected of being compromised by a poisoning attack that biases decisions against a specific demographic group. The bank must respond while maintaining regulatory compliance and minimizing customer impact. Which response sequence is most appropriate?

   • AImmediately retrain the model with the most recent data and redeploy it.
   • BShut down the model and manually review all credit applications indefinitely.
   • CIsolate the model, conduct a forensic analysis to confirm the attack, then retrain with clean data and validate the model before redeploying.
   • DAdjust the model's decision threshold to counteract the bias.
   Reveal Solution Discussion Select an answer first