Викторина, связанная с руководством · Средний Уровень

Редактирование PII в викторине LLM Pipelines

Сравните обнаружение и преобразование PII, обратимые токены, ложноотрицательные результаты, ограничения фильтра конфиденциальности OpenAI и почему редактирование не является анонимизацией.

Связанные направляющие путиРедактирование Pii в конвейерах LLM
Вопрос 1 из 8

What technique do PII redaction pipelines typically use to catch personal data with a predictable format, such as phone numbers or email addresses?

Предварительный просмотр каждого вопроса8 вопросы

Ответы и пояснения появляются по мере прохождения теста.

  1. What technique do PII redaction pipelines typically use to catch personal data with a predictable format, such as phone numbers or email addresses?
    • Regular expression pattern matching
    • Распознавание названного объекта
    • Homomorphic encryption
    • Federated learning
  2. Why do pipelines add a named entity recognition (NER) step in addition to regex matching?
    • To encrypt the detected entities
    • To detect personal names, addresses, and organizations that have no fixed format
    • To speed up regex matching
    • To generate the lookup table automatically
  3. In a table-based reversible-tokenization design, where should the mapping from placeholder tokens to original values be kept?
    • Inside the LLM's training data
    • Embedded directly in the token text
    • In a separate, access-controlled lookup kept in a trusted environment
    • It is not stored anywhere
  4. What risk does a reversible-token lookup table create?
    • It cannot be used to link placeholders
    • It makes every record anonymous
    • It removes the need to manage keys
    • It becomes a sensitive asset that needs access controls and protection
  5. Why does the guide say redaction is not the same as anonymization?
    • Because redacted text can sometimes still be re-identified from surrounding context clues
    • Because redaction always requires encryption keys
    • Because anonymization is illegal in most jurisdictions
    • Because redaction only works on numeric data
  6. Why can a stable pseudonym remain privacy-sensitive even when it replaces a direct identifier?
    • The original identity in plaintext
    • It can preserve links across records and may aid re-identification
    • The exact person’s real phone number
    • A guarantee that the replacement is safe to contact
  7. What should be checked after a redaction pass before content enters logs or a downstream model?
    • Only whether the prompt is shorter
    • Whether the language model can infer every placeholder
    • Whether serialization, errors, and downstream traces still contain sensitive spans
    • Nothing; one detector guarantees complete removal
  8. How does the guide characterize OpenAI Privacy Filter?
    • As a general compliance certification
    • As an API that automatically redacts all requests
    • As a reversible token vault for all cloud data
    • As an open-weight text PII detection and masking aid with known limits