Privacy-Preserving Federated Learning (PPFL) Testbed and Curated Dataset

Today, the NIST National Cybersecurity Center of Excellence (NCCoE) released the open-source Privacy-Preserving Federated Learning (PPFL) Testbed, along with a curated genomic dataset. The PPFL Testbed is a component of the NIST Privacy-Enhancing Technologies (PETs) Testbeds. Together, these resources give researchers and organizations a practical environment for testing how PETs can enable the use of sensitive data while protecting privacy.

Background

Machine learning (ML) models trained on sensitive and proprietary data can uncover new insights, for example in biomedical research, where patient data can help identify new treatments. However, organizations are often unable to share data because of privacy, regulatory, or security concerns.

PETs are technical controls that attempt to reduce privacy risks by limiting how data is collected, analyzed, and shared—which can enable secure collaboration. Federated learning lets organizations train a shared model while keeping data on site. Differential privacy adds calibrated noise during training to limit what the model reveals about any individual record, at some cost to accuracy. Genomic data is large, complex, and sensitive, which makes that balance even more challenging.

Through the PETs Testbeds, the NCCoE is developing open-source testbeds, datasets, and reference architectures to help organizations evaluate these technologies and accelerate their adoption.

Resources Available

To enable organizations to experiment with PETs, the project has released two resources:

• Open-Source Testbed: Open-source software for training ML models centrally or across federated sites, with or without differential privacy. This tool allows users to compare how experiment configurations and privacy settings affect model performance. The first use case is genomics, and the processing scripts in the companion data repository show how to prepare other genomic data. The code also works with many tabular datasets with little configuration.
• Curated Dataset: A machine learning-ready soybean genomic dataset covering over 18,000 accessions (or individual records) from the United States Department of Agriculture (USDA) Soybean Germplasm Collection. This serves as a common dataset to enable researchers to evaluate model performance across different testbed configurations.

This release builds on insights and collaboration from the broader genomics and privacy community, including a public red-teaming exercise in 2025.

For more details and information on how to use these resources, visit the PETs Testbeds page.

Stay Informed

To stay informed and get involved in the future development of PETs tools and resources, we encourage you to join the Genomic Data Community of Interest.  Please submit questions and feedback to [email protected].