A Strategy-Agnostic Framework for Partial Participation in Federated Learning

ICLR 2026 Conference Submission24959 Authors

Keywords: Partial participation, Stochastic optimization, Convex optimization, Non-convex optimization

Abstract: Partial participation (PP) is a fundamental paradigm in federated learning, where only a fraction of clients can be involved in each communication round. In recent years, a wide range of mechanisms for partial participation have been proposed. However, the effectiveness of a particular technique strongly depends on problem-specific characteristics, e.g. local data distributions. Consequently, achieving better performance requires a comprehensive search across a number of strategies. This observation highlights the necessity of a unified framework. In this paper, we address this challenge by introducing a general scheme that can be combined with almost any client selection strategy. We provide a unified theoretical analysis of our approach without relying on properties specific to individual heuristics. Furthermore, we extend it to settings with unstable client-server connections, thereby covering real-world scenarios in federated learning. We present empirical validation of our framework across a range of PP strategies on image classification tasks, employing modern architectures, such as FasterViT.

Primary Area: optimization

Submission Number: 24959