Is de-identification the same as encryption?

No. Encryption protects data using keys; authorized decryption restores the original information. De-identification changes or limits the identifying association in the data or output.

Can a de-identified dataset still be useful?

Yes, but utility depends on which details are preserved and the intended analysis.

Can de-identified data sometimes be re-identified?

Yes. Linkage, retained mappings and other information can restore associations in some circumstances.

AI Glossary / Graphs and symbolic AI / Intermediate

De-identification

Definition

De-identification is the process of reducing the association between data and the people or other subjects it describes. It can involve removing identifiers, transforming identifying combinations and controlling how data is released or accessed. It reduces disclosure risk rather than automatically eliminating it.

What it means

A name is an obvious identifier, but a record can remain recognizable after the name is removed. A rare job title, precise location and date may identify someone when combined with information available elsewhere.

De-identification therefore concerns both the dataset and its use context. A table shared inside a restricted research environment presents a different exposure from the same table published openly. The likely recipient's additional information also matters.

Terminology varies across disciplines and legal frameworks. Pseudonymization commonly replaces direct identifiers with substitute labels while retaining some ability to link records. This entry explains the technical process; it does not treat a particular transformation as sufficient to meet every jurisdiction's definition of anonymous data.

How it works

Begin by defining the intended analysis and sharing arrangement. Identify direct identifiers, such as names and account numbers, and quasi-identifiers: fields that can become identifying in combination. Check free text, attachments and metadata as well as obvious table columns.

Possible transformations include suppression, replacing precise values with broader groups, masking selected fields and using carefully designed synthetic or aggregate outputs. Pseudonyms can preserve links across records, but a retained mapping or predictable transformation may permit the association to be recovered.

Assess both disclosure risk and analytical usefulness after transformation. Coarsening location may protect against linkage but weaken geographic analysis; removing rare categories may hide the very cases the analysis needs. The chosen trade-off should be explicit.

The process also includes release controls and review. Repeated releases can reveal information when combined, and new external datasets can change the linkage risk. A tool that merely replaces names cannot by itself establish that the complete dataset and release plan are adequately de-identified.

Three practical examples

An analytics team removes direct identifiers and groups precise ages before sharing a customer-behavior dataset under a defined access arrangement.
A support-quality team reviews message text for identifying details that would remain after account names were removed.
A research service offers approved aggregate queries instead of releasing all individual records when that better fits the analysis and disclosure constraints.

Think of it this way

Think of removing name badges from a group photograph. People may still be recognized from their faces, uniforms or location. Removing the most obvious label does not remove every identifying clue.

The analogy is limited because structured records can be linked at scale using combinations that are hard to notice visually. De-identification needs a systematic assessment, not simply a judgment that individual fields look harmless.

Worked example

Why group size does not tell the whole story

Consider an invented table containing four records. Two have age band 30-39 and region North; two have age band 40-49 and region South. Names and precise ages have been removed.

With respect to the chosen pair of quasi-identifiers, every record shares its combination with one other record. The table therefore meets a simple k=2 indistinguishability condition for those two fields.

Now suppose both North records have the same sensitive outcome, “service suspended,” while the South records have different outcomes. Someone who already knows that a particular person belongs to the North group can infer that person's sensitive outcome even without identifying which of its two rows is theirs.

The example demonstrates an attribute-disclosure limitation. It does not estimate a real person's re-identification probability or establish that k=2 is an acceptable release threshold. Additional fields, background knowledge and the sharing environment could change the assessment further.

Illustrative case study

A fictional learning-platform company, Cedar Learning, prepares support tickets for an AI classification experiment. Its first export replaces user names with random labels but retains message text containing email signatures and precise descriptions of individual incidents.

The team inventories these indirect channels and revises the preparation workflow. It also separates the pseudonym mapping from the analysis copy and decides which recipients need record-level access.

Before using the transformed tickets, reviewers check a representative set for missed identifiers and assess whether the remaining text still supports the intended classification task. The company describes the result and residual limitations accurately. It does not label the data anonymous merely because one masking operation completed successfully.

Common misconceptions and mistakes

Removing names completes de-identification. Other fields and free text can still identify subjects.
Replacing identifiers with stable codes prevents all linkage. Stable pseudonyms intentionally preserve some links and may be reversible.
Meeting one table-based criterion eliminates all disclosure risks. Attribute inference and external information can still matter.
Synthetic data is automatically private. Its generation process and potential disclosure behavior need evaluation.

Business use

De-identification can enable useful analysis while reducing exposure of personal information. For AI projects, define whether the data supports development, evaluation, training or external sharing, because the access pattern and retained detail can differ.

Keep a record of the source data, transformations, intended uses, access conditions and assessment method. Evaluate the actual release artifact rather than assuming that every exported field followed the same treatment. Separate technical risk reduction from claims of legal status, and revisit the assessment when the data, recipients or use changes.

Check your understanding

A dataset replaces names with random IDs but keeps a free-text column containing names and email addresses. Has the replacement alone addressed the identifying content?

Answer: No. The free text still carries direct identifiers and must be included in the assessment and treatment.

Related terms

Anonymization
Synthetic data
Pseudonymisation
quasi-identifier
re-identification
data minimisation
disclosure risk

Learn next

Take an invented five-row table and mark direct identifiers, possible identifying combinations and sensitive attributes. Propose transformations for one specific analysis, then explain what useful detail would be lost and what linkage risks would remain.

Sources

NIST IR 8053, De-Identification of Personal Information, surveys techniques and re-identification limitations: https://csrc.nist.gov/pubs/ir/8053/final

NIST SP 800-188, De-Identifying Government Datasets: Techniques and Governance, describes risk assessment and sharing models: https://csrc.nist.gov/pubs/sp/800/188/final

The four-record example and Cedar Learning case are fictional teaching illustrations.

Previous: De Finetti's theorem Next: De Moivre's law