Threats and Vulnerabilities Specific to AI AAIA Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A security team wants to protect a computer vision model from adversarial inputs that cause misclassification. They have access to a large set of legitimate images and can generate adversarial examples. Which mitigation strategy would be most effective?

   • AData augmentation
   • BAdversarial training
   • CModel inversion
   • DInput validation
   Reveal Solution Discussion Select an answer first
2. 12foundation · easy

   Which of the following is a risk associated with using third-party pre-trained models in an AI system?

   • AThe model cannot be integrated with existing systems
   • BThe model may contain hidden backdoors or biases from its training data
   • CThe model will require more storage than a custom model
   • DThe model will always be slower than a custom-built model
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A medical research lab offers a model that predicts patient outcomes. An attacker queries the model with synthetic profiles and, by observing the confidence scores, is able to infer the approximate values of a specific patient's medical records that were part of the training data. Which AI-specific threat is this?

   • AData poisoning
   • BModel drift
   • CPrompt injection
   • DModel inversion
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A company deploys an LLM that has access to a database of customer records. An auditor is concerned about prompt injection attacks that could extract sensitive data. The company wants to minimize the risk while maintaining the LLM's functionality. Which control is most effective?

   • AUse a more powerful LLM with better safeguards
   • BRestrict the LLM's database access to only necessary fields
   • CImplement strict output filtering to block sensitive data patterns
   • DDisable the LLM's ability to process natural language
   Reveal Solution Discussion Select an answer first
6. 15foundation · easy

   Which technique is commonly used to craft adversarial examples that are imperceptible to humans but fool a model?

   • AEncrypting the input data
   • BChanging the model's activation functions
   • CFlooding the model with random inputs
   • DAdding small, carefully calculated perturbations to the input
   Reveal Solution Discussion Select an answer first