AI Assurance, Testing, and Auditing CRAGE Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A self-driving car company must test its perception model against adversarial conditions such as fog, rain, and sensor noise. Which type of AI testing is most appropriate for this scenario?

   • AFunctional testing with standard clear-weather inputs
   • BBias testing across demographic groups
   • CPerformance testing measuring inference speed
   • DRobustness testing with perturbed and adversarial inputs
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   Which technique is commonly used to provide local explanations for individual predictions of a machine learning model?

   • APrincipal Component Analysis (PCA)
   • BK-means clustering
   • CGradient descent
   • DSHAP (SHapley Additive exPlanations)
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   An AI audit is underway for a predictive policing system. The audit team has completed fieldwork and identified several high-severity issues, including potential bias and lack of transparency. The police department wants to release the audit report to the public. What should the audit team do before publication?

   • ADelay publication indefinitely until all issues are resolved.
   • BPublish the raw audit data and all findings without any redaction.
   • CRemove all negative findings to avoid public concern.
   • DVerify findings with additional evidence, and prepare a report that balances transparency with the need to protect sensitive operational details.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   An autonomous vehicle company is testing its perception model. The model performs well in sunny conditions but fails in fog. The team has a large dataset of sunny images but few fog images. Which testing methodology would be most effective in evaluating the model's robustness to fog?

   • APerformance testing to measure the model's inference speed and throughput.
   • BBias testing to compare the model's performance across different vehicle types.
   • CRobustness testing by applying synthetic fog effects to the sunny images and measuring the model's performance.
   • DFunctional testing using the existing sunny dataset to verify the model meets its specifications.
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A regulatory body requires a company to provide evidence that its AI system's decisions are interpretable. The company uses a gradient-boosted tree model. Which approach would provide the most rigorous evidence of interpretability for the regulator?

   • AProvide SHAP summary plots showing the global impact of each feature.
   • BUse a black-box explanation method that approximates the model with a linear regression.
   • CState that the model is based on decision trees, which are inherently interpretable.
   • DProvide the model's training code and hyperparameters.
   Reveal Solution Discussion Select an answer first