Test re-identification risk and real-image subgroup: Before approval, which evidence is most important?
Test re-identification risk and real-image subgroup performance because synthetic data may leak information and preserve imbalance.
The question
A manufacturer generates synthetic defect images to train a quality-inspection model. The vendor claims the images are anonymous, but some were derived from identifiable production photographs, and the synthetic set preserves the original defect imbalance. Before approval, which evidence is most important?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Measure model accuracy on synthetic images onlySynthetic-only testing may overstate performance and cannot establish privacy or representativeness in real production conditions.
- Compare synthetic and source-image distributionsDistribution comparison can reveal preserved imbalance, but does not by itself test whether identifiable information leaked into synthetic images.
- Increase synthetic-image volume for rare defectsMore synthetic examples may alter frequency, but do not establish anonymity or prevent inherited bias and source-information leakage.
- Test re-identification risk and real-image subgroup performance ✓Re-identification testing addresses privacy leakage, while real-image subgroup performance tests whether inherited imbalance harms deployment-relevant quality.
The trap
Treat synthetic data as requiring privacy and quality validation, especially when derived from identifiable or imbalanced sources. How to remember it
Test re-identification risk and real-image subgroup performance because synthetic data may leak information and preserve imbalance.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Development questions
- Conduct security testing alongside targeted robustness: Which plan is most defensible? →
- Adjudicate a stratified disagreement sample using: Which evidence most directly resolves that uncertainty? →
- Obtain access to linked model: Which evidence should the team obtain? →
- All 426 Understanding How to Govern AI Development questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Development ·
Every answer, right and wrong, comes with its own explanation.
