De-identification and synthetic data generation solve the same problem in fundamentally different ways. De-identification takes real patient records and subtracts information such as names, dates and ZIP codes, hoping that what remains cannot be traced back. Differentially-private synthesis generates entirely new records under a mathematical ceiling on how much any real patient can influence the output. The first is a policy whose strength depends on what attackers know. The second is a guarantee that holds no matter what they know.
How de-identification works, and why it keeps failing
Under HIPAA, de-identification takes one of two paths: Safe Harbor (remove 18 enumerated identifier types) or Expert Determination (a statistician certifies re-identification risk as "very small"). Both are subtraction: the records that remain are still real people's medical histories, minus the fields someone decided were identifying.
The failure mode is auxiliary information. Landmark work by Latanya Sweeney showed that 87% of Americans are uniquely identifiable by ZIP code, birth date, and sex alone. Those three fields routinely survive masking policies. Every subsequent decade added demonstrations: re-identification of "anonymous" movie ratings, genomic datasets, hospital discharge records, and location traces. The pattern is structural, not accidental: a masking policy defends against the auxiliary data its designer imagined. Attackers get to use auxiliary data the designer never imagined, including datasets that don't exist yet.
How differentially-private synthesis works
Instead of subtracting from real records, a differentially-private synthesis pipeline bounds how much any one person can influence what is learned from the source data. The literature describes two broad routes. One adds calibrated noise to statistics such as counts and marginals, then samples artificial records from those protected statistics. The other limits and noises each person's contribution while a generative model is trained. Either way, the privacy guarantee carries over to everything generated from the result, a property known as post-processing. Differential privacy guarantees that the presence or absence of any single patient changes the probability of any output by at most a bounded factor (eε). The bound is a theorem, not an assessment: it holds against any attacker, any auxiliary dataset, and any future technique, because it constrains the generation process itself rather than guessing what an adversary might know.
Side by side
| Property | De-identification / masking | DP synthesis |
|---|---|---|
| What ships | Real records, minus identifiers | Artificial records, statistically fitted |
| Privacy basis | Policy + expert judgment | Mathematical proof (ε, δ stated) |
| Auxiliary-data attacks | Principal failure mode | Covered by the guarantee |
| Future-proofness | Degrades as data accumulates | Bound is permanent |
| Statistical utility | Generalises or suppresses quasi-identifiers (dates, geography). Clinical detail, including rare codes, can stay largely intact, which is part of why it re-identifies | Preserves the schema and measured population structure, with fidelity reported. Rare tails are deliberately smoothed |
| Regulatory direction (EU) | Under increasing scrutiny | Judged by the EDPB draft's singling-out, linkability and inference tests; examined as a formal method in the TEHDAS2 access-body guideline (D7.2, March 2026) |
The regulatory trajectory
The EDPB's draft anonymisation guidelines (July 2026) treat synthetic data as anonymous only where identification is effectively precluded, a test we examine in Is synthetic health data anonymous under GDPR? In the US, HIPAA's Safe Harbor remains lawful, but the science it rests on ages every year, and expert determinations increasingly have to account for the modern auxiliary-data landscape. The direction of travel on both continents favours methods whose privacy can be demonstrated rather than asserted.
Honest caveat: the guarantee is not free. Differential privacy deliberately smooths distribution tails, which is why DP synthetic data is excellent for testing, development, feasibility, and ML benchmarking. It is genuinely unsuitable for rare-disease cohorts and safety-signal detection. A provider should tell you this before you ask. (Full suitability table: What Is Synthetic Health Data?)
Frequently asked questions
Is de-identified data still personal data?
Under the GDPR, the answer depends on whether individuals can still be singled out, linked, or inferred. Under HIPAA, properly de-identified data is no longer PHI. De-identified data can fail that test when auxiliary data enables re-identification. That gap between the two regimes is precisely where re-identification research has done its damage.
Why does de-identification destroy statistical utility?
The fields most useful for analysis, such as dates and geography, are also among the most identifying, so masking policies target them first. Generalising dates to years or truncating ZIP codes breaks the temporal and geographic joints that analytics depends on. Synthesis preserves the learned joint structure instead and reports the fidelity it achieved.
Does differential privacy mean the synthetic data is exactly as useful as real data?
No. The privacy budget (ε) is a dial between disclosure protection and statistical precision; small cohorts and rare categories lose the most. Trustworthy synthetic data ships with a measured fidelity report, so you know what was preserved rather than assuming it.
Can I combine both, de-identifying first and then synthesising?
Synthesis pipelines typically consume identified or pseudonymised source data inside the data holder's environment, so pre-masking mostly just degrades what the model can learn. The privacy guarantee should come from the synthesis mechanism, not from input masking.
- L. Sweeney, "Simple Demographics Often Identify People Uniquely" (Carnegie Mellon, 2000), the source of the 87% result
- HHS: HIPAA de-identification guidance, Safe Harbor and Expert Determination
- EDPB Draft Guidelines 02/2026 on Anonymisation (July 2026)
- TEHDAS2 Joint Action. D7.2 Guideline for Health Data Access Bodies on data minimisation, pseudonymisation, anonymisation and synthetic data. 24 March 2026. doi:10.5281/zenodo.20269109
- Synthetix Health Research: What Is Synthetic Health Data?
Continue reading in Research:
Go deeper
Our analysis pieces take the regulatory and practical questions this primer raises further.
Is synthetic health data anonymous under GDPR? Fit for which purpose?