Synthetic Data CompaniesGet a quote
Pages

Synthetic data privacy guarantees

A synthetic dataset is only as private as the method behind it. Differential privacy gives a measurable bound, set by a value called epsilon; de-identification and "no row is copied" checks reduce risk without proving anything. None of the vendor pages we read stated an epsilon.

Every vendor page cited here was last checked on .

What is the difference between differential privacy and de-identification?

How the main approaches compare. General descriptions, not claims about any vendor.
ApproachWhat it doesWhat you can claimCost to quality
MaskingKeeps real rows and hides or swaps sensitive valuesDirect identifiers are gone; the rest of each row is realLow
De-identified synthetic dataStrips identifiers, then trains a generator and samples new rowsNo row is a copy, if tests confirm it; no formal boundLow to moderate
Differentially private synthetic dataAdds calibrated noise during training so no single record shifts the model muchA stated bound (epsilon, delta) on what anyone can learn about one recordModerate to high, most of all for small tables and rare values

What does epsilon mean?

Epsilon is a privacy budget. It bounds how much the generator's output can change when any one record is added or removed. At small values, an attacker learns almost nothing about whether your record was there. As epsilon grows, the bound loosens quickly, because it works on an exponential scale.

Delta is the small probability that the bound fails. It should be far smaller than one divided by the number of records.

Three details decide whether a quoted epsilon means anything:

  • What is protected: a row, or a person who may have many rows. Per-row privacy is weaker when one customer has hundreds of transactions.
  • Where the budget is spent: training only, or also preprocessing such as choosing category lists and value ranges from the real data.
  • Whether it is total: each new release from the same source spends more budget.

A longer, plain-English walk-through is in differential privacy in synthetic data.

What does each vendor state about privacy?

From each vendor's own pages, read 7 Nov 2026. "Not published on the pages we read" means we found no statement we could quote, not that the vendor lacks one.

Privacy statements on vendors' own pages, read 7 Nov 2026
VendorData typesDeliveryPrivacy claim on its pages
Tonic.aiTables and textSoftware platformNot published on the pages we read
Syntho (and MOSTLY AI, powered by Syntho)TabularSoftware platformNot published on the pages we read
AindoTabularSoftware platformLists a Europrivacy certification (a GDPR certification scheme) on its security page
K2viewTabularSoftware platformNot published on the pages we read
SAS Data MakerTabularSoftware platformNot published on the pages we read
Rendered.aiImages, 3DPlatform plus servicesNot published on the pages we read
Parallel DomainImages, video, 3DSoftware platformNot published on the pages we read

Some vendors in this table may pay an introduction fee when a team we put in touch signs with them. Whether a vendor pays has no effect on the list, its order or our notes.

This site is independent. It is not affiliated with, sponsored by or endorsed by any company named on it, and it has not tested or audited any of their products: everything here comes from what each company publishes.

What should you get in writing?

  1. The privacy method, and if differential privacy, epsilon, delta and the unit protected.
  2. The privacy tests run on each delivered dataset and their results.
  3. Where your source data is processed and stored, and when it is deleted.
  4. Whether the trained generator is kept, and who may use it.
  5. What happens if a test later shows a real record was reproduced.
Where these facts come from

Privacy questions about synthetic data

Does synthetic data count as personal data?

It can. If a person can be singled out or inferred from the synthetic records, privacy law may still apply. Have your privacy counsel look at the method and the test results, not the label "synthetic".

What is a good epsilon for synthetic data?

There is no single safe number. Smaller is stronger. The value means little without knowing how it was spent across training, so ask the vendor for the total budget, delta, and what was protected (rows, or people with several rows).

Is differential privacy always worth it?

Not always. It costs fidelity, most of all on small tables and rare categories. For internal test data with no outside sharing, de-identification plus strong privacy tests may be a reasonable trade.

Which vendors offer differential privacy?

None of the vendor pages we read on 7 Nov 2026 stated an epsilon or a differential privacy guarantee in terms we could quote, so we print "not published". Ask each vendor in writing.

What is a membership inference attack?

A test of whether someone can tell, from the synthetic data or the generator, that a particular person's record was in the training set. A strong result for the attacker means the data exposes real people.

Get a quote

Three parts: the company location, a contact for replies, then what the generated data has to do. The final part is optional.

Disclosure: this is an introduction service, and any vendor it introduces you to can pay the site for that introduction. Each brief is read by a person and matched against vendors' published terms; you get the names first, and nothing goes to any vendor without your go-ahead.

Please do not send data, samples, schemas with real values or credentials through this form. A written description is all we need.

Part 1 of 3

Where is the company based?

Five digits are enough.

Get a quote