FENRIS: Federated Extensible Norwegian Research Interface for Synthetic data

Example Pipelines | CLI Reference | Extension Guide | Contributing | Changelog

FENRIS is an extensible benchmarking framework for federated synthetic tabular data generators. Built on Flower, it provides a reproducible pipeline covering dataset preparation, client partitioning, federated training, multi-dimensional evaluation, and artifact generation.

The framework is designed to be extended with new algorithms, partitioning strategies and evaluation metrics through a plugin system that allows new components to be added without modifying the core system. Algorithms are conceptualized as a combination of a synthesizer (the generative model) and a coordinator (the federated training strategy), allowing for flexible combinations and comparisons.

FENRIS ships with two literature-based reference implementations (Fed-TGAN and FedTabDiff) and evaluates algorithms across fidelity, utility, privacy, fairness, and scalability.

This framework was initially developed during Spring 2026 as part of a bachelor thesis at OsloMet in collaboration with the Norwegian Institute of Public Health (Folkehelseinstituttet, FHI) and the Cancer Registry of Norway (Kreftregisteret).

The following sections provide a quickstart guide to using the framework, including prerequisites, setting up the environment, verifying the installation, and running a benchmarking pipeline. For more detailed documentation and examples, please refer to the links at the top of this README.

Quickstart Guide

Prerequisites

FENRIS v0.1.0 supports Python 3.12 and 3.13. Linux and macOS are the recommended platforms; Windows support is on a best-effort basis due to constraints imposed by Flower's Ray backend. WSL may be used as a workaround.

Installation

The recommended installation method is pipx, which installs FENRIS into an isolated environment and exposes the fenris command globally without requiring a virtual environment:

pipx install fenris

To install pipx itself, follow the official installation instructions.

Alternatively, FENRIS can be installed into a virtual environment via pip:

python -m venv .venv
source .venv/bin/activate   # On Windows: .venv\Scripts\activate
pip install fenris

To verify a successful installation:

fenris --version

Acquiring the reference datasets

The reference datasets used in the examples are available at the following URLs, pinned to a stable commit:

• https://raw.githubusercontent.com/Gubbeklubben/FENRIS/6aac5c5/datasets/breast_cancer.csv
• https://raw.githubusercontent.com/Gubbeklubben/FENRIS/6aac5c5/datasets/heart_disease.csv
• https://raw.githubusercontent.com/Gubbeklubben/FENRIS/6aac5c5/datasets/lung_cancer.csv

Executing a pipeline

The following command runs a complete end-to-end benchmark pipeline on the Breast Cancer dataset using FedTabDiff, with the default seed of 42:

fenris run fedtabdiff fedavg iid_partitioner datasets/breast_cancer.csv

Once the run completes, FENRIS creates a timestamped output directory, for example:

out/2026-05-01-14-11-26-fedtabdiff-breast_cancer-<uuid>/

This directory contains the following artifacts:

A successful run produces non-null fidelity metric values in both metrics files. Utility, fairness, and attribute inference metrics require --target-col and/or --sensitive-cols to be specified and will evaluate to null otherwise, as in this minimal example.

Further documentation

Example Pipelines | CLI Reference | Extension Guide | Contributing | Changelog