NoisyFlow

NoisyFlow is a three-stage pipeline for federated synthetic data generation with optional differential privacy. It trains a flow-matching generator per client, fits an optimal transport map to a target domain, and then synthesizes target-like data for downstream classification tasks. The framework has two main goals: (1) enable domain adaptation by generating data that matches a target distribution, so that classifiers trained on the synthetic (domain-transferred) data perform better on the target domain than classifiers trained only on the original (source) domains; and (2) allow combining both original and transferred data to further improve classification accuracy on the desired target domain (3) all of these have privacy guarantee.

Features

• Stage 1: flow matching generator with optional DP-SGD (Opacus).
• Stage 2: ICNN or CellOT transport (options A/B/C).
• Stage 3: server-side synthesis and classifier training.
• Optional privacy-utility sweeps and membership inference evaluations.

Installation

python -m venv .venv
source .venv/bin/activate
pip install torch numpy
pip install pyyaml  # required for YAML configs
pip install opacus  # optional, DP-SGD in stage1/stage2
pip install matplotlib  # optional, privacy curve plots

Quickstart 22

python run.py --config configs/default.yaml

For a smaller smoke test:

python run.py --config configs/quick_smoke.yaml

Note: configs/quick_smoke.yaml sets device: cuda; switch to cpu if you do not have a GPU.

To run the toy demo script:

python noisyflow_sketch.py

Configuration

• Configs live under configs/ and are loaded by run.py.
• data.type supports federated_mixture_gaussians, mixture_gaussians, and toy_federated_gaussians.
• Enabling DP (stage1.dp / stage2.dp) requires Opacus; stage2 DP requires stage2.cellot.enabled: true and stage2.option: A.
• privacy_curve.enabled: true runs a sweep and writes privacy_utility.png (requires matplotlib).

Documentation

Start here: docs/README.md.

• docs/overview.md: Pipeline overview and stage summary.
• docs/configuration.md: Full config reference.
• docs/data.md: Synthetic data builders.
• docs/experiments.md: CLI usage and experiment configs.
• docs/attacks.md: Membership inference attack details.
• docs/architecture.md: Code map.

Tests

python -m unittest

Tests that depend on optional packages (pyyaml, opacus, matplotlib) are skipped if those packages are not installed.