Skip to content
Technology transformation expertise firm · ARDIANE Group Innovate, secure, succeed

Data

Data anonymisation: techniques, limits and choices according to use

In brief

Data anonymisation makes it impossible, by reasonable means, to identify the people in a dataset, which then falls outside the scope of the GDPR; a pseudonymised dataset remains personal data for anyone who can trace back to the person. This Management Square article compares the techniques, indicates which one to choose according to use and describes how to verify the result.

Page contents

Analytics, testing and AI teams ask for realistic data. The data protection officer rightly points out that the production database contains personal data. Data anonymisation is then presented as the solution, even though the word covers very different operations.

A truly anonymous dataset falls outside the scope of the GDPR; a pseudonymised dataset remains within it for anyone who holds, or can reasonably obtain, the means of tracing back to the person. No technique guarantees anonymity on its own: the result is assessed dataset by dataset. Finally, the right level of protection depends on the use: many needs are better served by openly acknowledged pseudonymisation, in a controlled environment, than by approximate anonymisation.

These pointers do not constitute legal advice: the classification of a dataset belongs to the controller, advised by its data protection officer.

Anonymisation or pseudonymisation: what the classification changes

The GDPR, Regulation (EU) 2016/679, defines pseudonymisation in Article 4, point 5: the data can no longer be attributed to a person without the use of additional information, kept separately and protected by technical and organisational measures. As long as a mapping table or a key makes it possible to trace back to the person, the data remains personal for whoever holds it or can reasonably access it.

Recital 26 of the same regulation states that the principles of data protection should not apply to anonymous information, that is information which does not or no longer allows a person to be identified. It gives the criterion: account should be taken of all the means reasonably likely to be used to identify someone, by the controller or by any other person, having regard to the cost, the time required and the available technology.

Pseudonymised dataset Anonymous dataset
Status Personal data for anyone who can reasonably re-identify Outside the scope of the GDPR
Applicable framework The GDPR in full (legal basis, purpose, retention, rights, security) No further GDPR obligations as long as anonymity holds; other protections may remain, for example the confidentiality of communications
What is kept The link between records relating to the same person Trends, sometimes with a heavy loss of granularity

Pseudonymisation is not a second-rate measure: Article 32 of the GDPR cites it among security measures. It reduces risk without changing the status of the data for whoever retains the means of re-identification.

The boundary between the two notions is still the subject of case law and of work by the European Data Protection Board, particularly for a recipient that holds no means of re-identification. When in doubt, a pseudonymised dataset is treated as personal data.

Data anonymisation techniques and what they do not guarantee

Opinion 05/2014 of the Article 29 Working Party on anonymisation techniques, which predates the GDPR and has been adopted by the CNIL, distinguishes two main families.

Generalisation dilutes attributes: a year instead of a date of birth, a department instead of an address. Aggregation and k-anonymity belong to it: each combination of indirectly identifying attributes must be shared by at least k people. Its limit: if all the members of a group carry the same sensitive value, belonging to the group is enough to reveal it. The l-diversity and t-closeness variants address this flaw.

Randomisation alters values: adding noise, swapping between records. Differential privacy is its most formalised form: calibrated noise bounds how much a person’s presence changes the result of a computation. Its limit: protection depends on the level of noise chosen and erodes as queries accumulate, while noise that is too strong strips the analysis of meaning.

Synthetic data, produced by a model trained on the real dataset, is not covered by this opinion and belongs to neither family. Its limit: an overly faithful generator can reproduce real records or recognisable rare cases.

Hashing, tokenisation and encryption, for their part, belong to pseudonymisation. Removing direct identifiers is not enough either: a postcode, a date of birth and an occupation, combined, can designate a single person.

Choosing according to use: testing, analytics, model training

Test and acceptance datasets. The need concerns format, consistency between tables and coverage of edge cases; the truth of the values matters little. Fictitious data, generated from rules, is often enough and avoids the question. When a production extract is necessary, it is most often a matter of pseudonymisation: the test environment must then offer the protections of production (restricted access, traceability, retention period).

Analytics and management reporting. Management indicators are calculated on aggregates. Generalisation is suitable, provided a minimum group size is set and cross-tabulations are monitored: two tables that are harmless separately can, by difference, isolate an individual. For analyses of individual journeys, anonymity is hard to achieve, because a detailed trajectory is often unique: pseudonymisation in a controlled workspace is then the honest answer.

Model training. A model needs the granularity that anonymisation removes. Two statuses need examining: that of the training dataset and that of the model, which can memorise and then reproduce elements of its data and therefore cannot be presumed anonymous. The levers: reducing attributes to what is necessary, pseudonymising, testing for memorisation, using differential privacy or verified synthetic data. For high-risk systems, Article 10 of the AI Act, Regulation (EU) 2024/1689, subjects training, validation and testing datasets to data governance practices. Its Article 4a, inserted by Regulation (EU) 2026/1744, allows the processing of special categories of data to detect and correct bias only exceptionally, under several cumulative conditions, including that this cannot be achieved with other data, including synthetic or anonymised data. Under amended Article 113, Article 10 applies from 2 December 2027 for Annex III uses and from 2 August 2028 for systems linked to Annex I products.

This choice is best made once per data category and per use, then applied in the processing of the data platform, rather than renegotiated at every extraction.

Checking that a dataset is really anonymous

Opinion 05/2014 proposes three criteria, adopted by the CNIL, which can be posed as three questions:

  1. Singling out: can an individual be isolated in the dataset?
  2. Linkability: can records concerning the same person be linked, within this dataset or with another?
  3. Inference: can new information about a person be deduced with a high degree of probability (the CNIL writes “almost certainly”)?

When the three criteria are not all met, the CNIL requires an in-depth assessment of re-identification risks. Verification brings together four elements:

  • Measurements on the dataset: share of unique records on combinations of indirectly identifying attributes, size of the smallest group.
  • A re-identification attempt, carried out by a team that did not produce the dataset, using the sources a third party could draw on (public data, other datasets already released).
  • An examination of the context: to whom the dataset is given, with which other data, for how long.
  • A file and a review date. Re-identification techniques and available sources evolve, and the CNIL requires regular monitoring: a dataset that is anonymous today may no longer be so tomorrow.

Uniqueness and group size measurements can be automated in the same pipelines as data quality checks.

Why a dataset described as anonymous often is not

  • Labelling a pseudonymised dataset “anonymised”. The label spreads through the catalogue, and the dataset circulates without protection.
  • Reasoning column by column. The risk arises from combinations of attributes, not from an isolated field.
  • Forgetting free-text fields and technical identifiers. Comments, logs, IP addresses and device identifiers often carry what the structured columns no longer contain.
  • Judging the dataset in isolation. Successive versions of the same dataset, or extracts given to different teams, can be cross-checked.
  • Believing the matter is closed after the operation. Anonymisation is itself processing of personal data, and the source dataset remains subject to the GDPR.

Anonymity is demonstrated dataset by dataset

Data anonymisation never results from merely applying a technique. Three actions structure it: classifying each dataset honestly, choosing the protection according to the use rather than the tool available, and verifying and then reviewing the result over time.

None of them holds without identified owners: an owner who decides whether a dataset is opened, a data protection officer consulted on its classification, a classification that follows it in the catalogue. Our data governance consultants help put these roles in place and translate these rules into access policies and effective controls.

FAQ

Frequently asked questions

Another question? Ask us

What is the difference between anonymisation and pseudonymisation?

Pseudonymisation replaces identifying elements with a pseudonym, but information kept separately (a mapping table, a key) makes it possible to trace back to the person: for whoever holds that information or can reasonably obtain it, the data remains personal within the meaning of the GDPR, Regulation (EU) 2016/679. Anonymisation aims at a dataset in which nobody can any longer be identified by reasonable means. Only this second result falls outside the scope of the regulation.

Is anonymised data still subject to the GDPR?

Recital 26 of the GDPR states that the regulation does not apply to anonymous information. The anonymity must still be real: it is assessed in the light of all the means reasonably likely to be used to identify a person. The anonymisation operation itself is carried out on personal data, and the source dataset remains subject to the regulation.

Is hashing or encryption enough to anonymise data?

No. Opinion 05/2014 of the Article 29 Working Party classifies hashing, encryption and tokenisation among pseudonymisation techniques. As long as a key, a mapping table or a simple recalculation of the hash makes it possible to find the person, the data remains personal for whoever has that means. These techniques reduce the risk; they do not turn a pseudonymised dataset into an anonymous one.

Is synthetic data anonymous?

Not by nature. A synthetic dataset is produced by a model trained on real data, and an overly faithful generator can reproduce existing records or recognisable rare cases. A synthetic dataset is therefore verified like any other dataset that claims to be anonymous, before any release.

How do you check that a dataset is really anonymous?

European data protection authorities use three criteria from Opinion 05/2014 of the Article 29 Working Party, adopted by the CNIL: singling out, linkability and inference. Verification combines measurements on the dataset, a re-identification attempt by a separate team and an examination of the release context. It is documented and repeated over time.

A project to bring to completion?

A first conversation is enough to understand your challenge, identify the right people and tell you frankly what we can bring.