↓Skip to main content

Rashomon Sets and Model Multiplicity in Federated Learning

·2 mins

TL;DR #

Rashomon Sets and Model Multiplicity in Federated Learning studies what happens when several Federated models perform almost equally well in terms of privacy but differently in terms of other metrics such as fairness.

Problem #

Federated Learning usually produces one global model by averaging updates from many clients. That model can hide important differences between clients: a model that performs well on average may still fail on a particular client or group of clients. In centralized Machine Learning, the collection of similarly accurate but behaviorally different models is known as a Rashomon set. In this paper we studied how to define and study those sets without collecting clients’ raw data in the Federated Learning setting.

What did we do? #

We formalized three federated perspectives:

  • Global Rashomon set: models that satisfy a performance constraint across aggregated client data.
  • t-agreement Rashomon set: models that satisfy the constraint for a chosen fraction of clients.
  • Individual Rashomon set: models that each client considers sufficiently accurate for its own data.

We also designed a privacy-compatible pipeline for generating candidate models. The pipeline requires at most two additional communication rounds beyond standard training.

Main findings #

Experiments on Dutch Census, ACS Income, and MNIST show that global evaluation can hide large differences in local model behavior. Increasing the client-agreement threshold makes the set smaller and filters out models that only perform well on average. Individual clients can also observe multiplicity patterns that are very different from global averages.

Resources #

Cite this paper #

@inbook{10.1145/3805689.3812290,
author = {Heilmann, Xenia and Corbucci, Luca and Cerrato, Mattia},
title = {Rashomon Sets and Model Multiplicity in Federated Learning},
year = {2026},
isbn = {9798400725968},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3805689.3812290},
booktitle = {Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency},
pages = {2434–2461},
numpages = {28}
}