Privacy-Preserving Federated Learning – Future Collaboration and Continued Research – Source:www.nist.gov

Reflections and Wider Considerations

This is the final post in the series that began with reflections and learnings from the first US-UK collaboration working with Privacy Enhancing Technologies (PETs). Since the PETs Prize Challenges, the ecosystem around these technologies has continued to develop, with a shift from more theoretical and academic conversations to greater uptake and consideration of PETs.

Since our first post in December of 2023, this series has explored a variety of practical considerations relevant to working with Privacy-Preserving Federated Learning (PPFL), ranging from understanding different types of privacy attacks and ways to mitigate their risks to individuals’ privacy, to exploring the importance of input and output privacy. Through recent posts, we also heard from guest contributors who were winners and judges in the PETs Prize Challenges.  They shared thoughts and considerations for scalability, implementation, data pipeline challenges, and more.

The scope of this blog series reflects the breadth of insights and considerations that emerged from the PETs Prize Challenges, but there are further aspects of working with PPFL that the challenges – and this series – have not addressed. This includes considerations for working with real data across multiple jurisdictions. PPFL can remove the need to transfer, store, or process real data centrally; this is useful when data sets must remain decentralized (for policy or technical reasons). However, we need research to inform our understanding of real-world applications of PPFL and PETs in general. NIST is researching the considerations and nuisances of real-world PPFL deployments through the PETs Testbed, hosted by the NIST National Cybersecurity Center of Excellence (NCCoE).

Future Collaboration

As a relatively novel and emerging approach to working with data in a privacy-centric way, PPFL has the potential to support greater innovation and foster collaboration in the future.

The UK and US have agreed to further collaborative initiatives on PETs. Building on this commitment, and on insights from our previous collaboration, the UK National Disease Registration Service and the US National Cancer Institute are working together over the next few months to explore how PETs can drive innovative research into rare pediatric cancers through secure, privacy-preserving data collaboration between national disease registries. By using PETs, researchers can do cross-border data analysis without the need for data transfer or direct access; they can gain deeper insights without compromising data privacy. This activity is being supported by the UK’s Department for Science, Innovation and Technology, the White House’s Office for Science and Technology Policy, NIST, and the US Department of Energy, and coordinated with the National Science Foundation.

Collaboration using PETs, such as PPFL, will enable researchers to work around challenges that arise from the scarcity of data in individual countries. Using a federated approach as a mechanism for querying and/or modeling data will allow researchers to analyze data on rare pediatric cancers in a privacy-preserving way. This will also allow researchers to analyze information in ways that were previously not possible due to the limited availability of data (at present, no single jurisdiction has access to a large enough dataset encompassing ultra-rare tumor types to conduct such analysis on their own). 

This approach also offers potential to scale research in the longer term, to include additional data from other jurisdictions in the analysis. Scaling to more countries could support a broader global initiative to foster collaboration on pediatric cancer.

Continued Research

The wider ecosystem around PPFL continues to develop, with the establishment of more forums for discussion and more opportunities for policymakers and researchers to align and collaborate on the horizon.

To further investigate PETs and their respective suitability for specific use cases, NIST has launched the PETs Testbed. In collaboration with the U.S. Census Bureau XD team and offered through the NCCoE, the first model problem is a PPFL model architecture with a genomics use-case. The model solution will allow us to explore the contours of the difficulty in deploying PPFL to solve real-world problems. NIST is creating privacy and utility metrology to support understanding of their relationship (e.g., tradeoffs) in the context of federated learning. The architecture will undergo a privacy threat evaluation that will involve the use of tools such as the NIST Privacy Framework and the outcomes of a privacy red-teaming exercise. These assessments will serve as a framework for helping organizations navigate the trade-offs in a PPFL system.

Additional information

As an emerging approach for working with data, there is still much to learn and explore about PPFL, and PETs more broadly, in theory and practice. The links below provide further information on this, including examples of real-world use cases:

• NIST PETs Testbed
• PETs Prize Challenges
• Costs and benefits associated with PETs
• OECD information and resources on PETs 

If you would like to share feedback or additional ideas, please reach out to us at pets [at] dsit.gov.uk (pets[at]dsit[dot]gov[dot]uk) or privacyeng [at] nist.gov (privacyeng[at]nist[dot]gov).