Skip to main content

​
Models in VeriSynth

VeriSynth uses machine learning models to learn the statistical structure of real data and generate synthetic data that behaves like the real world — without exposing sensitive records. Each model is designed to balance realism, privacy, and computational efficiency.
The system is modular, so users can plug in new synthesis engines over time.

​
Current Model: Gaussian Copula (GC)

The Gaussian Copula Synthesizer is the first and default model in VeriSynth Core.
It’s a fast, deterministic, and lightweight model for tabular synthetic data — perfect for most structured datasets.

​
How It Works

  1. Learn relationships
    GC learns pairwise correlations between all columns in your dataset.
    It transforms them into a continuous Gaussian space where relationships are easier to model.
  2. Model dependencies
    A multivariate Gaussian distribution is fit to the transformed data.
    This captures how each variable depends on others (e.g., age ↔ BMI ↔ blood pressure).
  3. Sample synthetic data
    New synthetic samples are drawn from this learned distribution, then mapped back to the original data space.
  4. Output verification
    The generated dataset is post-processed and evaluated for correlation deltas, privacy risk, and consistency metrics.

​
Example

Let’s say your real dataset has these correlations:
Variable PairCorrelation
Age ↔ BMI0.42
BMI ↔ Systolic BP0.61
Age ↔ Glucose0.58
The synthetic dataset produced by VeriSynth will preserve these relationships closely, often within a correlation delta of ±0.1 — enough to be statistically realistic for most analytical and ML tasks.
Instead of random noise, GC learns your dataset’s statistical DNA and regenerates it faithfully.

​
Strengths

FeatureDescription
LightweightRuns entirely on CPU — no GPUs required
DeterministicSame seed → same result, ideal for reproducibility
ExplainableTransparent mathematical model (no black box)
FastGenerates up to 1M+ rows in seconds
AuditableOutputs proof receipts with correlation metrics

​
Limitations

LimitationDescription
Linear correlations onlyGC struggles with nonlinear or multimodal relationships
Limited categorical complexityHigh-cardinality categories may be oversimplified
No temporal or sequential modelingBest for static, tabular datasets

​
Roadmap: Upcoming Models

VeriSynth Core is model-agnostic — future releases will expand into deep generative and privacy-enhanced models.
ModelTypeStatusDescription
CTGANGAN-based tabular🧩 PlannedCaptures nonlinear relationships and rare categories
TVAEVariational Autoencoder🧩 PlannedProduces smoother synthetic distributions and uncertainty estimates
DP-GCDifferentially Private Gaussian Copula🧩 PlannedAdds formal privacy bounds (ε) to correlation modeling
TimeSynthSequential/time-series🧩 PlannedModels synthetic patient timelines, transactions, or sensor data
ImageSynthGenerative Vision🚧 ExploratoryExtends proof-based synthesis to visual data

​
Model Selection Philosophy

We believe every synthetic data generator should be:
  1. Understandable — transparent about how it learns and samples
  2. Reproducible — deterministic seeds, reproducible metrics
  3. Verifiable — accompanied by measurable fidelity and privacy proofs
  4. Modular — easy to swap models as new techniques evolve
VeriSynth is built as a modular framework, so you can run: 
verisynth data/patients.csv -o out/ --model gaussian_copula
# (future)
verisynth data/patients.csv -o out/ --model ctgan

​
Model Registry (Coming Soon)

We’re working on a model registry system that will let users:
  • View available models (verisynth models list)
  • Inspect metadata, dependencies, and required hardware
  • Register custom synthesis engines via plugin
  • Benchmark models on fidelity vs privacy
Example: 
verisynth models list
# → gaussian_copula, ctgan, tvae, dp_gc

​
Model Validation Metrics

Each model is evaluated using:
MetricDescription
Correlation delta (Δ)Measures difference in variable relationships
KS test (p-value)Checks distributional similarity
Naive re-identification riskApproximates privacy exposure
Synthetic utilityPerformance of ML models trained on synthetic data
These metrics are logged to proof.json for transparency and reproducibility.

​
Example CLI Run (GC Model)

verisynth data/finance.csv -o out/ --rows 50000 --model gaussian_copula --seed 42
Produces: 
📁 out/synthetic.csv
🧾 out/proof.json
With proof: 
{
  "model": { "engine": "gaussian_copula", "seed": 42 },
  "metrics": { "corr_mean_abs_delta": 0.15, "naive_reid_risk": 0.0 }
}

​
Summary

PrincipleDescription
Start simpleGC gives accurate, transparent tabular synthesis
Build modularlyNew models can plug into the same proof pipeline
Stay verifiableEvery model, no matter how complex, will always produce a proof receipt