합성 데이터 퀴즈

퀴즈를 풀면 정답과 설명이 나타납니다.

1. A set of synthetic records is designed to test an empty form field. What does that use establish?
   • The records exercise a chosen interface case, not the real-world frequency of that case
   • The proportion of real users with empty fields
   • That the generator preserves every population relationship
   • That the records are private by construction
2. Why keep generated and observed records distinguishable in lineage?
   • To make synthetic records look more realistic to end users
   • To document how examples were produced and what information influenced them
   • To ensure generated examples always match real data
   • To avoid evaluating the generator
3. A synthetic set looks plausible overall but omits rare failure conditions. What is the concern?
   • Plausible data guarantee subgroup coverage
   • Rare cases are always irrelevant
   • It may fail to represent cases that matter for the downstream task
   • Overall appearance proves statistical validity
4. What does a privacy guarantee tell you about a synthetic set’s usefulness?
   • That it is suitable for every analysis
   • That it preserves every correlation
   • That it reflects the real population exactly
   • Nothing by itself; privacy and task utility require separate evaluation
5. A generator was trained on sensitive records. Which privacy assumption is unsafe?
   • That generated data cannot reveal source information merely because identifiers are absent
   • That the generation method affects privacy risk
   • That privacy claims need a threat model
   • That a mechanism’s parameters matter
6. Which evaluation best matches synthetic data intended to support model development?
   • Check only whether synthetic rows look realistic
   • Compare task performance against an appropriate independent real-world test set
   • Use the generator’s training score as the downstream result
   • Assume matching column averages is sufficient
7. A team claims differential privacy from a synthetic-data pipeline. What should reviewers inspect?
   • Only the number of generated rows
   • Whether names were removed
   • The actual privacy mechanism and parameters behind the claim
   • The font used in the data dictionary
8. Why might a dataset useful for interface testing be unsuitable for population statistics?
   • Interface tests always contain too many columns
   • Synthetic rows cannot be counted
   • Population statistics require images
   • Its cases may have been deliberately selected for edge coverage rather than sampled representatively