Category: Privacy & Data Protection

Re-Identification Risk

Also known as: re-identification risk analysis

Simply put

Re-identification risk is the likelihood that data which has been anonymized or de-identified can be linked back to the specific individuals it describes. This typically happens when the de-identified dataset is combined with other available information, allowing a third party to figure out who the data subjects are. Even data that appears anonymous can carry some level of this risk.

Formal definition

Re-identification risk is the likelihood that a third party can re-identify data subjects within a de-identified or anonymized dataset, commonly by linking residual data elements to external information sources. As commonly defined, assessment approaches examine the data properties and record attributes that overlap with externally available data to estimate the probability that individual subjects can be re-identified. Note that the term describes a residual privacy risk associated with de-identification rather than a guarantee of anonymity; its treatment and acceptable thresholds vary by sector (for example, healthcare and research contexts) and by applicable jurisdictional requirements, which are out of scope for this core definition.

Why it matters

Re-identification risk matters because de-identification and anonymization are frequently treated as endpoints that make data safe to share, reuse, or release, when in practice they leave a residual privacy risk. As commonly defined, this risk arises when de-identified records retain data elements that overlap with externally available information, allowing a third party to link records back to specific individuals. Organizations that assume anonymization is absolute may under-protect data that can still be traced to the people it describes, exposing data subjects to privacy harm and the organization to regulatory and reputational consequences.

The stakes are particularly pronounced in sectors such as healthcare and research, where sensitive information about individuals is routinely de-identified for secondary use. In these contexts, acceptable re-identification risk thresholds and the treatment of residual risk vary by applicable jurisdictional requirements and sector-specific expectations. Because these thresholds are not universal, teams cannot rely on a single standard of anonymity across all uses; what is considered adequately de-identified in one setting or jurisdiction may not satisfy another.

Understanding re-identification risk as a measurable, residual property—rather than a binary state of anonymous versus identifiable—helps organizations make defensible decisions about data sharing, publication, and reuse. It reframes de-identification as a risk-reduction measure that lowers, but does not necessarily eliminate, the possibility of subjects being re-identified.

Who it's relevant to

Data Privacy and Protection Officers

Privacy professionals rely on re-identification risk assessment to determine whether data has been de-identified sufficiently for a given use, sharing arrangement, or release. Understanding that de-identification leaves a residual risk—rather than producing absolute anonymity—is essential to making defensible decisions and to selecting acceptable risk thresholds appropriate to context and jurisdiction.

Healthcare and Clinical Data Teams

Teams handling patient information frequently de-identify data for secondary use, where re-identification risk is the chance that anonymized patient information can be traced back to individuals. Because acceptable thresholds and treatment vary by sector and jurisdiction, these teams need to assess residual risk explicitly rather than assuming de-identified records are fully anonymous.

Researchers and Data Stewards

Those preparing research datasets for sharing or publication use re-identification risk frameworks to identify data elements that overlap with external data and to estimate the probability that subjects could be re-identified. This supports informed decisions about how datasets are released and what additional protections may be warranted.

Data Scientists and Engineers

Practitioners who build and manage data pipelines can conduct re-identification risk analysis to find properties in sensitive data that might increase the risk of subjects being identified. This helps them apply de-identification as a risk-reduction measure and communicate residual risk to privacy and compliance stakeholders rather than presenting anonymization as complete.

Inside Re-Identification Risk

Re-Identification

The process by which an individual is linked back to data that has been de-identified, anonymized, or pseudonymized, defeating the intended privacy protection. Re-identification risk is the likelihood that such linkage can occur given available data, techniques, and auxiliary information.

Quasi-Identifiers

Attributes that are not unique identifiers on their own but, in combination, can single out an individual (for example, ZIP code, date of birth, and sex). They are a central driver of re-identification risk and are frequently the focus of risk assessment.

Auxiliary or Background Information

External datasets or knowledge that an adversary can combine with the released data to re-identify individuals. Because such information evolves over time, re-identification risk is not static and can increase after data release.

Linkage Attacks

A common re-identification technique in which a de-identified dataset is joined to another dataset sharing overlapping attributes to match records to identified individuals.

De-Identification and Anonymization Techniques

Methods intended to reduce re-identification risk, such as generalization, suppression, aggregation, and perturbation. These reduce rather than eliminate risk, and their effectiveness depends on the dataset and threat context.

Adversary Model / Threat Assumptions

The assumptions about who might attempt re-identification, their motivation, and the resources and auxiliary data available to them. Risk estimates are meaningful only relative to a stated adversary model.

Risk Metrics

Quantitative or qualitative measures used to estimate re-identification likelihood, such as measures based on the uniqueness or population size of records sharing a combination of quasi-identifiers. The choice of metric affects the resulting risk assessment.

Common questions

Answers to the questions practitioners most commonly ask about Re-Identification Risk.

Does anonymizing or de-identifying data eliminate re-identification risk?

No. De-identification typically reduces re-identification risk rather than eliminating it. Even after removing direct identifiers, individuals may be re-identified through combinations of quasi-identifiers, linkage with external datasets, or inference from model outputs. As commonly framed, re-identification risk should be treated as a residual risk to be measured and managed, not a hazard that is fully removed by a single anonymization step.

Is re-identification risk only relevant when direct identifiers like names or IDs are present in the data?

No. A common error is to assume that removing obvious identifiers makes data safe. Re-identification frequently arises from quasi-identifiers—attributes that are not uniquely identifying on their own but become so in combination—and from linkage to other available data. The presence or absence of direct identifiers is only one factor; the broader context, including auxiliary data and release model, typically matters as much or more.

How can an organization assess the level of re-identification risk in a dataset before release?

Assessment approaches commonly combine analysis of quasi-identifiers, evaluation of how unique records are within the dataset, and consideration of what external data an adversary might plausibly access. The assessment typically also accounts for the release context—for example, whether data is shared publicly, with a limited set of parties, or held internally—since the same dataset can carry different risk depending on who can access it and what they can link it to.

What role does the threat model or assumed adversary play in evaluating re-identification risk?

The assumed adversary is central to any evaluation. Re-identification risk is generally not an intrinsic property of the data alone but depends on the capabilities, motivation, and auxiliary information available to a potential attacker. Practitioners commonly document explicit assumptions about the adversary and the release environment so that the risk assessment can be interpreted, challenged, and revisited as those assumptions change.

How should re-identification risk be monitored after data or a model has been deployed?

Because re-identification risk can change over time as new external datasets and linkage techniques become available, it is commonly treated as something to review periodically rather than assess once. Monitoring practices may include re-evaluating risk when new auxiliary data sources emerge, when access arrangements change, or when model outputs could reveal information about individuals, and updating controls accordingly.

How does re-identification risk relate to broader AI governance and model risk management activities?

Re-identification risk can surface within both AI governance—through policies, accountability, and oversight of how data is handled—and model risk management, where it may be considered among the risks arising from model use and outputs. These are related but distinct: governance addresses the organizational structures and controls, while model risk management focuses on identifying, measuring, monitoring, and controlling the risk itself. Where the two overlap, the risk should be addressed without collapsing the distinction between them.

Common misconceptions

Anonymized or de-identified data carries no re-identification risk and therefore falls outside privacy obligations.

De-identification typically reduces but does not necessarily eliminate re-identification risk. Residual risk can remain and may change over time as auxiliary information becomes available, so the applicability of privacy obligations depends on the specific technique, dataset, and legal framework in question.

Removing direct identifiers such as name and Social Security number is sufficient to prevent re-identification.

Quasi-identifiers that remain after direct identifiers are removed can, in combination, single out individuals, particularly when combined with external datasets through linkage attacks. Direct identifier removal is one step rather than a complete safeguard.

Re-identification risk is a fixed property of a dataset that can be measured once and considered settled.

Risk is relative to the assumed adversary, the available auxiliary information, and the techniques applied, all of which can evolve. A dataset assessed as low risk at release may present higher risk later as new external data or methods emerge.

Best practices

Define an explicit adversary model and threat assumptions before assessing re-identification risk, so that risk estimates are interpreted against a stated set of resources and auxiliary data rather than in the abstract.

Identify and evaluate quasi-identifiers in combination, not only direct identifiers, since combinations of attributes are a primary driver of linkage-based re-identification.

Treat de-identification and anonymization techniques as risk-reducing measures rather than guarantees, and document the residual risk that remains after applying them.

Reassess re-identification risk periodically and when circumstances change, recognizing that newly available auxiliary information can increase risk after initial release.

Select and document risk metrics appropriate to the dataset and use case, noting how the choice of metric influences the assessment and its limitations.

Consider controls beyond the data itself, such as access restrictions and release conditions, since technical de-identification alone may not adequately manage residual re-identification risk.