Skip to main content

Frequently Asked Questions (FAQ)

⚠️ Work in Progress
This page is under active development. Information may be incomplete or subject to change.

This section gathers answers to the most frequently asked questions about the Avatar anonymization method. Anonymization is a nuanced process influenced by legal context, dataset structure, and intended use. Below, we clarify key points and best practices to help you get the most out of the tool.

❓ Can we revert to the original individual from an anonymized dataset?​

No, reversing an anonymized record to retrieve the original individual is not possible.
To qualify as anonymization, a process must be irreversible under legal standards.
To achieve this, the Avatar method introduces stochastic variation during the local modeling of each synthetic record. See more in the methodology section.

❓ Can the method handle small datasets (a few dozen records)?​

Yes, the method supports small datasets — but with a few caveats:

  • A small number of individuals increases the risk of reidentification due to uniqueness.
    At the same privacy level, fewer statistical properties can be preserved compared to a larger dataset.

  • Avatar-generated data reflects the confidence interval of the source data.
    With fewer records, this interval is wider, meaning results may vary more across iterations.

❓ Does the method support free text anonymization?​

Not directly. Free text is unstructured and, as of now, lacks a clear legal framework to define what constitutes anonymous textual data.

However, it is possible to structure the text using NLP, anonymize the structured version with Avatar, and then regenerate text (if needed) using a language model.
This ensures personal data isn’t used to train any model.
We have successfully implemented this workflow in past projects — contact us to learn more.

❓ Can images be anonymized using this method?​

Not at this stage. There is currently no clear legal definition for what makes an image anonymous.

However, you can anonymize structured data linked to pseudonymized images, then later reassociate them via probabilistic matching.
We’ve used this framework in real-world use cases — reach out for more details.

❓ Does anonymization require user consent?​

Not necessarily.
When anonymization is done properly and demonstrably, it’s considered compatible with the original purpose of data collection under GDPR.

The Avatar method automatically provides two complementary reports (technical assessment and impact assessment) that can be shared with your DPO.
Therefore, explicit user consent is not required for anonymization using Avatar.

❓ How do we assess whether a means is reasonable?​

The EDPB does not require a means to be impossible in the abstract. It asks whether a means is reasonably likely to be used in the real world, taking into account cost, time, technical feasibility, and the actual environment of disclosure.

The assessment must therefore consider several practical points:

  • Chain of means: a recipient does not need to be able to execute the attack himself. If he has the technical or financial ability to call on a specialist or third party to do it, that is a reasonable means.
  • Multiplicity of actors: the analysis is not limited to the legitimate recipient. It must include disloyal colleagues, neighbors or relatives, investigative journalists, public authorities, questionable companies, and cybercriminals.
  • No “no motivation” assumption: the EDPB rejects the idea that anonymity can be concluded just because a potential actor has “no interest” in re-identifying someone.
  • Illegal means are not automatically irrelevant: compliance by legitimate actors can only be presumed. One cannot conclude that a means is unreasonable simply because it may be unlawful.
  • Contractual clauses are not enough: a contractual prohibition on re-identification is useful, but it is only a complement to technical and organizational safeguards and never sufficient by itself to establish anonymity.

In short, the question is not whether a hostile actor would be willing to invest effort, but whether such effort is realistically possible and proportionate to the value of the information. This is the central logic of the “reasonable means” test.

For the official source, see the EDPB guidance on anonymisation: EDPB Guidelines 02/2026 on anonymisation.

❓ Can I anonymize my entire database at once for multiple use cases?​

Yes, the tool supports anonymizing an entire database in one go. A database run reads a database, anonymizes every group of linked tables, and writes the result into another database.
However, we recommend splitting your anonymization projects by use case whenever possible.

This enables you to fine-tune the privacy/utility balance for each specific purpose, resulting in better-quality avatars.
Learn more in the step-by-step guide.

❓ Does the method support wide matrices?​

Yes — the method can handle datasets with more variables than records.
However, this configuration increases reidentification risk, as each record becomes highly unique.

At equal privacy levels, fewer statistical properties can be preserved in wide matrices than in datasets with more individuals than variables.

❓ Does the method reduce dataset variance?​

Slightly, yes.
To protect unique records, the method automatically centers individuals with rare values, leading to a small reduction in variance.

However, weak signals — like those in small clusters or rare events — are preserved, ensuring the dataset remains usable for analysis and modeling.

❓ Can certain results be preserved exactly in the anonymized dataset?​

Yes, the method can be configured to preserve specific values exactly.
This naturally impacts privacy, but that impact is quantified by the metrics in the report, helping you decide whether the trade-off is reasonable for your context.

❓ What level of bias does the method introduce?​

Results from synthetic data typically fall within the confidence interval of the original data.

  • Less structured signals (e.g., noisy or less robust patterns) tend to introduce more variability.
  • Well-structured and stable signals are preserved almost identically.

The method involves randomness, so multiple iterations may be performed to get closer to your key indicators — especially when you know what to target in advance.

❓ Why was a patient-centric model chosen for anonymization?​

The patient-centric approach was chosen to ensure transparency, explainability, and control. Unlike generative AI models, this design lets data scientists fully understand and guide the anonymization process, reducing risks such as model misbehavior or hidden reidentification leaks. It also makes it easier to explain how data is protected to individuals. This model further enables unique, individual-level privacy metrics not possible in aggregate approaches.

❓ Can the method retain the characteristics of small clusters of individuals?​

Yes — the method can preserve rare but non-unique patterns in the data. Unique individuals are automatically recentered to prevent reidentification risk, which also helps clean the dataset. However, if a small group shares a rare trait, you can preserve that cluster by adjusting the anonymization parameter k to be lower than the group size. This makes the method suitable for analyzing rare events or subpopulations without sacrificing privacy.

❓ When is my data considered anonymous?​

Data anonymity is not a binary status — it depends on the context of use, the sensitivity of the data, and the realism of the attack scenarios that could be used against it. To assess this, you must evaluate the residual re-identification risk using criteria defined by the EDPB. The Avatar method supports both the simplified approach and the contextual approach described in official EDPB guidance.

The simplified approach is the conservative, worst-case view used for strict-sharing scenarios such as open data or broad external distribution. The contextual approach is the case-specific assessment used when the actual recipient, purpose, environment, and safeguards are known and controlled.

The Avatar method provides automatic risk metrics and recommended thresholds for strict use cases like open data. These metrics are consolidated into a technical report that quantifies the residual risk and helps document the assessment.

Learn more in the Privacy Metrics section

In addition, a PIA (Privacy Impact Assessment) is provided to evaluate the likelihood of potential attacks against the anonymization process, as well as their impact if successful. In simple terms, the technical report evaluates the residual re-identification risk posed by certain threats, while the PIA measures the probability and severity of those threats in the given context.

The recent EDPB guidance reinforces that a dataset can only be considered anonymous for a given use case when the residual risk remains sufficiently low and the technical and organizational safeguards are proportionate to the risk. Importantly, this is not a “100% protection on every metric” test: the question is whether the residual risk is acceptable in context, and whether the chosen safeguards and assumptions remain proportionate to the realistic attack model.

Where residual risk is considered too high, additional technical and organizational safeguards can be implemented to further strengthen the anonymization process.

❓ Can this method be used for data augmentation?​

We developed parameters that allow you to generate synthetic datasets larger than the original while preserving statistical properties. You can either augment the size of your dataset or rebalance modalities for underrepresented groups. We also calculate specific privacy metrics for augmented data to ensure that reidentification risks remain acceptable. Learn more in the data augmentation section.

❓ How does the method handle missing data?​

The method natively supports missing data and can preserve the same missingness patterns (MAR, MNAR, etc.) in the synthetic dataset. This is useful when the absence of data itself carries meaning. Alternatively, it is also possible to generate a version of the synthetic data without any missing values if needed for downstream tasks.

❓ What projection models are used in the anonymization process?​

Projection models are selected automatically based on your dataset:

  • PCA for numeric data
  • MCA for categorical data
  • FAMD for mixed datasets
  • FPCA or Fourier transforms for time series

All chosen methods are reversible to suit the methodology described for the avatar approach. This allow the output synthetic data to have the same structural and encoding properties as the original data. Non-reversible projection methods like t-SNE or UMAP can't be used for this reason.

❓ How do you handle computed variables and business rules?​

We recommend excluding computed variables (e.g., BMI) from the anonymization process only when both the computed variable and its input components (e.g., height and weight) are present in the dataset. In such cases, it's preferable to anonymize only the components and recalculate the derived variable afterward. This ensures logical consistency and reduces redundancy. Additionally, the tool includes processors to enforce business rules, such as sequential dates or fixed-sum constraints across variables.

❓ Who is responsible in case of improper data anonymisation?​

As a data processor, we provide an anonymisation tool along with a technical assessment of residual re-identification risks. This enables the data controller (i.e. the user of the tool) to make an informed decision based on the specific context of data use.

Under the General Data Protection Regulation (GDPR), the data controller remains ultimately responsible for ensuring the lawfulness and effectiveness of anonymisation. The technical report supports that decision, but the legal accountability remains with the controller.

See Article 24 GDPR – Responsibility of the controller

💡 More questions coming soon — and feel free to contact us if your specific case isn’t covered here.

  • Does the method support genomic data?
  • Data Characteristics and Impact
  • Are there examples comparing original and anonymized datasets?
  • How can I anonymize a data stream?
  • How do you handle anonymization in relational databases?
  • How does the method anonymize time series data?
  • Can GPS traces be anonymized?
  • How do you handle computed variables and business rules?