A data scientist investigates reports that a production machine learning (ML) model no longer performs...

Basic Concept: When unauthorized changes to an ML training pipeline cause model degradation, the root cause is insufficient access control and change management around the pipeline. Preventing future occurrences requires implementing governance controls that ensure all pipeline changes are reviewed and approved before execution. CompTIA SecAI+ Study Guide covers MDLC change management controls.

Why B is Correct: Enabling human review and approval workflows in the repository creates a mandatory gate requiring authorized reviewers to examine and approve any changes to training pipeline code before they can be merged and executed. This prevents unauthorized modifications from reaching the pipeline by enforcing a review process where any suspicious or unauthorized changes will be caught and rejected before they affect model training and performance.

Why A is Wrong: Static code scanning analyzes code for vulnerabilities and coding standard violations. While it improves code quality and security, it does not prevent unauthorized individuals from submitting and merging malicious changes to the pipeline without proper review.

Why C is Wrong: Retraining with more data and epochs addresses model performance restoration after the fact but does not prevent future unauthorized pipeline modifications. If the pipeline remains unprotected, the same attack could occur again on the new model.

Why D is Wrong: Keeping multiple model copies enables rapid restoration of a previous version when a deployed model is found to be compromised. While useful for recovery, it is a reactive measure that does not prevent unauthorized pipeline changes from occurring and affecting future model training.