Research in Collaborative Learning Does Not Serve Cross-Silo Federated Learning in Practice

TL;DR

This paper investigates why cross-silo federated learning (FL) remains underutilized in practice by conducting semi-structured interviews with 21 stakeholders across data owners, platform providers, and researchers. It reveals a substantial misalignment between the literature’s focus on technical optimization and practitioners’ concerns about awareness, trust, governance, and regulatory complexity. The authors organize findings along the FL pipeline—ideation, prototyping, evaluation, and deployment—highlighting concrete barriers at each stage and providing concrete directions for future work. The work argues for realigning FL research priorities toward end-to-end deployment, industry standards, and usable tooling to enable real-world cross-organizational collaboration.

Abstract

Cross-silo federated learning (FL) is a promising approach to enable cross-organization collaboration in machine learning model development without directly sharing private data. Despite growing organizational interest driven by data protection regulations such as GDPR and HIPAA, the adoption of cross-silo FL remains limited in practice. In this paper, we conduct an interview study to understand the practical challenges associated with cross-silo FL adoption. With interviews spanning a diverse set of stakeholders such as user organizations, software providers, and academic researchers, we uncover various barriers, from concerns about model performance to questions of incentives and trust between participating organizations. Our study shows that cross-silo FL faces a set of challenges that have yet to be well-captured by existing research in the area and are quite distinct from other forms of federated learning such as cross-device FL. We end with a discussion on future research directions that can help overcome these challenges.

Figures (3)

• Figure 1: Misalignment of challenges studied in Cross-silo FL research vs. those discussed in interviews. We compare the frequency of 12 selected challenges from analyzing the top 200 cited papers on "cross-silo federated learning" and our 21 interviews with FL stakeholders. Challenges are grouped by domain (model, data, systems, privacy, incentives, general). Literature focuses heavily on technical challenges such as model communication and data heterogeneity (left), while practitioners emphasize organizational barriers such as awareness and platform needs (right), revealing a fundamental disconnect between academic priorities and real-world deployment challenges. See Appendix \ref{['appendix:interview_plot']} for details on how this plot was generated.
• Figure 2: An example diagram of a cross-silo FL training procedure. Each client is an organization with its own data silo. At each communication round, (1) clients perform local training, and (2) upload the trained model to the server. The server then (3) aggregates the client models into a global model. Lastly, clients (4) download the new aggregate model and repeat this process.
• Figure 3: We compare the number of stakeholders who mentioned each of our 6 main topics during the interview. We find that each stakeholder group emphasizes different topics (relative to other stakeholders): clients care about models and data while platforms care about systems, privacy, and standards. Percentages are normalized by both the stakeholder group size and total times the topic is mentioned i.e. equal percentages mean an equal fraction of each stakeholder group, not an equal number of stakeholders.