Our philosophy
We believe sharing data to train AI should not expose sensitive information. SynthGAN generates plausible data products that preserve global statistical properties without reproducing real individuals.
Privacy is not an add-on: it is the starting point of synthetic design.
Security by design
Non-specific data, rarity diagnostics and privacy metrics integrated in every run.
Continuous validation
Real vs synthetic comparison: distributions, correlations, logical rules and consistency.
Interoperability
CSV, Parquet and JSON metadata exports ready for catalogs, APIs and data spaces.
What to expect from the lab
The SynthGAN lab is the project's public technical demo. Upload a reference CSV or use demo datasets and get a full synthetic generation report.
You get
- Automatic profiling of variable types and normality.
- Synthetic sample with correlations, partial correlations or betas.
- HTML report with charts, diagnostics and privacy metrics.
- Reproducible exports with fixed seed and traceable metadata.
It is not
- A copy of real records or a differential privacy guarantee.
- A substitute for clinical, regulatory or legal validation.
- A production SaaS product with authentication and SLA.
- An automatic decision on data use without human oversight.
How SynthGAN works
Three stages from reference CSV to validated synthetic data product.
-
Input
Reference real CSV, sample configuration, seed and dependency method.
-
Processing
Profiling, rule discovery, multivariate latent modeling, transformations and sampling.
-
Output
Synthetic dataset, validation report, exports and metadata for data space integration.
synthmulti pipeline
Architecture of the statistical engine powering the lab.
Define schema
Variable types, marginal distributions and constraints inferred from the CSV.
Latent model
Multivariate normal with correlations, partial correlations or betas per chosen parameterization.
Transform and sample
Continuous, binary, ordinal and categorical variables with post-process logical rules.
Diagnose and export
Statistical comparison, privacy, versioned metadata and interoperable files.
Who it is for
| Profile | Value |
|---|---|
| Health and longevity | Synthetic cohorts for research without exposing real clinical data. |
| Public administration | Policy simulation and planning with secure aggregated data. |
| Data spaces | Interoperable products with metadata for European catalogs and hubs. |
| AI teams | Training, QA and validation with non-sensitive reproducible datasets. |
Documentation and resources
Technical project materials. Demo datasets are synthetic or public domain; they contain no sensitive information and must not be used for automated decisions without validation.
SynthGAN technical sheet
Engine capabilities, lab usage and usage warnings.
API documentation — synthmulti MVP
Public library contract: schema, generator, diagnostics and invariants.
Demo dataset — mixed profiles
Demonstration CSV with continuous, binary, ordinal and categorical variables.
Demo dataset — Iris
Classic Fisher Iris dataset for quick pipeline tests.
Usage limits
Commitment to rigor
- Transparency: every report documents parameters, seed and dependency method.
- Traceability: JSON metadata with generator version and schema hash.
- No substitution: synthetic data supports analysis; it does not replace expert judgment.
- Generated datasets are demonstrative. They do not constitute legal, clinical or regulatory advice.