ISACA Advanced in AI Audit - AAIA Free Exam Questions

A. Differential privacy applied during model training

B. Supervised learning with labeled patient records

C. Data augmentation during training to improve privacy

D. Transfer learning using public health data sets

A. A large training dataset

B. High model accuracy

C. A well-written user manual

D. Version control of code, data, and model parameters

A. Optimize the decision threshold for maximum F1 on validation.

B. Increase the capacity of the model with early stopping.

C. Tune decision thresholds using expected misclassification costs.

D. Use class-weighted loss with stratified train/validation splits.

A. The electrical grid uptime is expected to be 72% of the time.

B. Grid failure is predicted to occur after 90 days.

C. The AI model correctly predicts maintenance needs 95% of the time.

D. F1 indicates that the model identifies true maintenance needs 76% of the time.

A. Sensitivity and origin of training data

B. Ability to generate diverse outputs

C. Frequency of model updates

D. Cleaning and validation methods for training data

A. As part of the initial design

B. During data collection

C. During the testing phase

D. Prior to moving into production

A. Establish responsibility and clear terms for model updates and support.

B. Only use models from vendors with globally recognized accreditation.

C. Ensure a minimum of three quotes have been obtained for market research and comparison.

D. Use the vendor only if the contract has been reviewed by the information security department.

A. Efficiency of reviewing procedures for compliance with regulations

B. Identifying outliers indicating data errors affecting model predictions

C. Processing cost associated with data analytics

D. Ability to utilize sample data population for model predictions

A. Shadow data

B. Copyright-free data

C. Extended retention of internal data

D. Data from web scraping

A. Unlabeled data sets

B. Labeled data sets

C. Clustered data sets

D. Randomized data sets

A. Translate all outputs to English and evaluate in English for consistency

B. Human review for non-English languages after go-live

C. Weighted metric with higher accuracy targets

D. Comparative language evaluations with parity thresholds