Skip to main content

Detection vs De-identification

Guardian Layer separates privacy workflows into two related steps:

  1. Detection finds sensitive content.
  2. De-identification changes that content to produce safer output.

Most applications can go directly to de-identification. Use detection first when you need to inspect, review, or control what will be changed.

Detection

Detection identifies sensitive terms and returns metadata. It does not change the input text.

Use detection when you need to:

  • Review what Guardian Layer found before changing the text.
  • Build a human review UI.
  • Audit sensitive terms without creating a de-identified copy.
  • Decide whether a document or message needs de-identification.

Endpoint:

POST /api/v1/analyze/text/proprietary

Example request:

{
"text": "The Meridian account is managed by Dana Reyes.",
"domain": "General",
"pca_dimension": 2
}

Response fields include:

FieldMeaning
tokensTokens recognized in the proprietary vocabulary.
sensitive_wordsSensitive terms detected in the input.
words_in_hullRelated terms found by semantic analysis.
suggestionsCandidate replacements for detected terms.
pca_dimensionThe PCA dimension used for the analysis.

De-identification

De-identification creates a safer output by changing detected sensitive values.

Use de-identification when you need to:

  • Send safer text to an LLM.
  • Share de-identified content with another system.
  • Generate privacy-preserving sample data.
  • Prepare content for downstream processing.

For Guardian Layer text de-identification, use:

POST /api/v1/deidentify/text/proprietary/outputs

Set masking_type based on the output style you want:

masking_typeBehavior
transformReplaces sensitive values with readable alternatives.
redactReplaces sensitive values with masking marks.

For detailed request and response examples, see Transform (Proprietary) and Redact / Masking / Placeholder.

Choosing a Workflow

GoalRecommended workflow
See what Guardian Layer would detectUse detection.
Produce de-identified text immediatelyUse de-identification.
Let a human approve changes firstDetect first, then de-identify after review.
Produce readable de-identified textUse masking_type: "transform".
Hide sensitive values more visiblyUse masking_type: "redact".

Next Steps