引導式測驗 · 硬 等級
ML 訓練管線中的 PII 處理測驗
減少攝取、轉換、日誌和工件中的 PII 暴露,同時了解偵測限制、假名化和存取控制。
問題 1 的 8
Which practice minimizes PII collection in an ML workflow?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Which practice minimizes PII collection in an ML workflow?
- Copy every collected attribute into a training table for possible future use
- Hash account IDs but retain every unrelated sensitive field
- Collect only personal data needed for the stated task
- Keep email addresses as model features even when the task does not use them
- Why can PII detection miss identifying information?
- Every detector only works on numeric fields
- Formats and contextual mentions vary across text, images, and metadata
- PII always follows one regular expression
- Detectors guarantee perfect recall across modalities
- How does pseudonymization differ from irreversible de-identification?
- A reversible token is anonymous because its mapping is encrypted
- Hashing an identifier always makes it impossible to relink records
- Removing a name means no party can recover identity
- A protected mapping can reconnect pseudonyms to people
- Why can removing direct names fail to eliminate re-identification risk?
- Combinations of attributes may identify someone
- Removing names eliminates all risk of identification
- Generalizing one field guarantees the remaining data is anonymous
- Identifying information occurs only in explicitly named columns
- Which data should be considered when scanning an ML workflow for PII?
- Only the original source table needs inspection
- Transformed features and model outputs cannot contain personal information
- Raw data, notebooks, logs, checkpoints, and shared artifacts
- Only public endpoint responses can retain examples
- What does least-privilege access mean for sensitive data?
- Give each workflow a shared administrator credential
- Give each user or job only the access needed for its task
- Assign every job the same broad workspace role
- Allow access based only on who created the dataset
- What privacy risk comes from logging raw user payloads?
- Logs automatically inherit deletion schedules from raw data stores
- Monitoring systems cannot retain image or audio request payloads
- Model outputs never contain information found in requests
- Sensitive data can persist in logs and be copied into downstream systems
- What should happen after direct identifiers are redacted?
- Assess residual linkability and model output risks
- Assume the dataset is anonymous once names are removed
- Remove access controls from the transformed dataset
- Store the re-identification mapping beside the shared dataset