Conditioning on Local Statistics for Scalable Heterogeneous Federated Learning
Rickard Brännvall
TL;DR
The paper addresses heterogeneity in federated learning by conditioning models on locally computed statistics, such as means, covariances, and higher moments, computed privately on each client. It introduces a framework where local statistics $$mu_i$$ are fed into models (e.g., conditional linear, ensemble, MLP, and a CNN variant for EMNIST) alongside the client’s data, enabling prediction that respects local distribution shifts without increasing communication. Empirical results on synthetic tasks and EMNIST show that these locally conditioned approaches achieve near-cluster-oracle performance and outperform global or client-only baselines, while remaining scalable and privacy-preserving. The work suggests that local-statistics conditioning offers a practical, communication-efficient mechanism to tailor federated models to heterogeneous data distributions and points to compression of statistics (e.g., via PCA) and extension to other modalities as future directions.
Abstract
Federated learning is a distributed machine learning approach where multiple clients collaboratively train a model without sharing their local data, which contributes to preserving privacy. A challenge in federated learning is managing heterogeneous data distributions across clients, which can hinder model convergence and performance due to the need for the global model to generalize well across diverse local datasets. We propose to use local characteristic statistics, by which we mean some statistical properties calculated independently by each client using only their local training dataset. These statistics, such as means, covariances, and higher moments, are used to capture the characteristics of the local data distribution. They are not shared with other clients or a central node. During training, these local statistics help the model learn how to condition on the local data distribution, and during inference, they guide the client's predictions. Our experiments show that this approach allows for efficient handling of heterogeneous data across the federation, has favorable scaling compared to approaches that directly try to identify peer nodes that share distribution characteristics, and maintains privacy as no additional information needs to be communicated.