SynthHub

SynthHub is a dataframe-first Python library for differentially private synthetic data. It gives DP synthesizers a small, scikit-learn-like interface:

from synthhub import Synthesizer

synth = Synthesizer(method="privbayes", epsilon=1.0, random_state=0)
synth.fit(real_df)
synth_df = synth.sample(1000)
report = synth.evaluate(real_df, synth_df, target="label")

The goal is not to invent another synthetic-data algorithm. SynthHub wraps existing DP engines behind one API, then makes preprocessing, evaluation, and privacy accounting visible in a common report.

Contents

• Why SynthHub
• Project Status
• Install
• Quickstart
• Backends
• Benchmark
• Privacy Contract
• Development
• Roadmap

Why SynthHub

DP synthetic-data tooling is fragmented. Popular synthetic-data libraries often separate privacy guarantees from their default open-source workflow, while research-oriented DP implementations expose different data formats, schema assumptions, and accounting conventions.

SynthHub focuses on the integration layer:

• one pandas DataFrame API for fitting, sampling, and evaluation
• backend adapters for DP synthesizers such as PrivBayes, AIM, MST, MWEM, and PATE/DP GAN families
• explicit PrivacyReport output for requested epsilon, spent epsilon, delta, backend identity, and warnings
• utility and privacy audit metrics for comparing methods under the same dataset and epsilon

Where SynthHub fits:

Project	Best at	SynthHub difference
SDV	Broad synthetic-data product surface	DP-first open adapter layer with explicit accounting reports.
SmartNoise Synth	DP mechanisms and synthesizers	Dataframe-first wrapper, common preprocessing, and shared evaluation.
SynthCity	Large research-oriented synthetic-data suite	Smaller API focused on DP backend switching and comparable reports.
DataSynthesizer	Classic DP PrivBayes implementation	Modern package shell, tests, benchmark, and unified Synthesizer API.

Project Status

SynthHub is in alpha. The public API is intentionally small, but backend coverage and evaluation reports are still evolving.

What is solid today:

• core Synthesizer.fit/sample/evaluate flow
• schema inference and explicit public schema support
• DataSynthesizer PrivBayes live smoke coverage
• Private-PGM AIM/MST live smoke coverage with upstream mechanisms
• SmartNoise MWEM/AIM/MST live smoke coverage
• adapter contract tests for optional backends
• CI across Python 3.10, 3.11, and 3.12 for the core package
• PyPI package release, build checks, and wheel install smoke coverage

What is still experimental:

• SmartNoise GAN and SynthCity adapters are contract-tested but not yet live-tested in CI
• Private-PGM AIM/MST require external mechanism modules on PYTHONPATH
• membership-inference scoring is an audit heuristic, not a DP proof

Install

Install from PyPI:

python -m pip install "synthhub[datasynthesizer]"

Install the current development version from GitHub:

python -m pip install "synthhub[datasynthesizer] @ git+https://github.com/tauptlab/synthhub.git"

Optional backend families:

python -m pip install "synthhub[smartnoise]"
python -m pip install "synthhub[synthcity]"
python -m pip install "synthhub[private-pgm]"

Private-PGM AIM/MST also require the upstream mechanisms/ folder on PYTHONPATH; see docs/private-pgm.md.

For local development:

git clone https://github.com/tauptlab/synthhub.git
cd synthhub
python -m pip install -e ".[test,datasynthesizer]"
python -m pytest -q

Quickstart

import pandas as pd
from synthhub import Synthesizer

real_df = pd.DataFrame(
    {
        "age": [21, 34, 37, 45, 52, 23, 41, 29, 62, 31],
        "city": ["A", "B", "A", "C", "B", "A", "C", "C", "B", "A"],
        "churn": [0, 1, 0, 1, 1, 0, 1, 0, 1, 0],
    }
)

synth = Synthesizer(method="privbayes", epsilon=1.0, random_state=0)
synth.fit(real_df)

synth_df = synth.sample(100)
report = synth.evaluate(real_df, synth_df, target="churn")

print(synth.privacy_report_.to_dict())
print(report.to_dict())

Try the notebook version in examples/quickstart.ipynb or open it directly in Colab.

Backends

Method	Backend family	Install extra	CI status	Notes
privbayes	DataSynthesizer correlated mode	datasynthesizer	live smoke	Default practical DP backend today.
datasynthesizer-privbayes	DataSynthesizer correlated mode	datasynthesizer	live smoke	Explicit alias for privbayes.
datasynthesizer-independent	DataSynthesizer independent mode	datasynthesizer	live smoke	Useful baseline over independent attributes.
independent	SynthHub one-way marginals	none	live tests	Built-in smoke-test baseline; not a production synthesizer.
aim	Private-PGM AIM	private-pgm plus mechanisms path	live smoke	Requires Private-PGM mechanism modules.
mst	Private-PGM MST	private-pgm plus mechanisms path	live smoke	Requires Private-PGM mechanism modules.
mwem	SmartNoise Synthesizers	smartnoise	live smoke	Epsilon-only mechanism; unsupported delta is handled explicitly.
pacsynth	SmartNoise Synthesizers	smartnoise	adapter contract	Optional dependency is heavy.
dpctgan, patectgan, pategan, dpgan	SmartNoise Synthesizers	smartnoise	adapter contract	Experimental GAN-family adapters.
smartnoise-aim, smartnoise-mst	SmartNoise Synthesizers	smartnoise	live smoke	SmartNoise-specific AIM/MST aliases.
synthcity-privbayes, synthcity-pategan, synthcity-dpgan	SynthCity privacy plugins	synthcity	adapter contract	Experimental until live CI is added.

Missing optional backends fail closed with BackendNotAvailableError and an installation hint.

Benchmark

The benchmark is reproducible and network-free:

python benchmarks/run_benchmark.py

It runs sklearn classification and regression datasets at epsilon=1.0 across installed backends. Full outputs are committed in benchmarks/results/latest.md and benchmarks/results/latest.csv.

Dataset	Method	Backend	Epsilon spent	Utility similarity	TSTR score	Re-ID risk
breast_cancer	independent	SynthHub baseline	1.000	0.788	0.636	0.086
breast_cancer	privbayes	DataSynthesizer correlated	1.000	0.598	0.715	0.036
breast_cancer	datasynthesizer-independent	DataSynthesizer independent	1.000	0.831	0.378	0.030
iris	independent	SynthHub baseline	1.000	0.784	0.307	0.000
iris	privbayes	DataSynthesizer correlated	1.000	0.739	0.480	0.000
iris	datasynthesizer-independent	DataSynthesizer independent	1.000	0.762	0.167	0.000
diabetes	independent	SynthHub baseline	1.000	0.776	-0.418	0.000
diabetes	privbayes	DataSynthesizer correlated	1.000	0.654	0.081	0.000
diabetes	datasynthesizer-independent	DataSynthesizer independent	1.000	0.772	-0.055	0.000

Utility similarity is a per-column distribution-similarity score. TSTR score is train-on-synthetic, test-on-real accuracy. Re-ID risk is a nearest-neighbor membership-inference heuristic and is not a DP proof.

Privacy Contract

Formal DP guarantees are conditional on public preprocessing metadata. This includes column names, column types, categorical domains, numeric bounds, and binning choices. If SynthHub infers these from private data, the PrivacyReport includes a warning.

For formal usage, start with docs/public-schema.md and pass an explicit public Schema to Synthesizer.

SynthHub verifies adapter-level contracts:

• requested epsilon is passed to the backend
• reported epsilon_spent does not exceed requested epsilon
• backend identity and accountant source are recorded
• output columns match the fitted dataframe schema
• DP-disabled modes such as SmartNoise disabled_dp=True are rejected

Read docs/dp-guarantees.md for backend-specific accounting sources, caveats, and CI coverage.

Development

Common checks:

python -m pytest -q
python -m pip wheel . --no-deps -w dist
python benchmarks/run_benchmark.py

Repository files:

• docs/design.md: architecture and adapter boundaries
• docs/dp-guarantees.md: privacy contract details
• docs/private-pgm.md: Private-PGM AIM/MST setup
• docs/public-schema.md: formal public-schema usage
• docs/release.md: PyPI release process
• benchmarks/run_benchmark.py: public benchmark
• CONTRIBUTING.md: development and PR expectations
• SECURITY.md: privacy/security reporting policy

Roadmap

Near-term:

• configure tokenless PyPI Trusted Publishing for future releases
• add richer benchmark datasets and normalized benchmark history
• add live CI coverage for one GAN-family backend

Later:

• OpenDP contingency-table adapter
• multi-table synthesis API
• CLI benchmark runner
• richer privacy attacks and utility reports

Community

Issues and feature requests are welcome. Please use private reporting for security-sensitive or privacy-accounting issues. Contributions should follow CONTRIBUTING.md.