Global Pivots, Local Unknowns: Stable Federated Open-Set Semi-Supervised Learning

You Rim Choi, Seojun Heo, Subeom Park, Jaehee Lee, Eunchung Noh, Hyung-Sin Kim

Keywords: Federated Learning, Federated Semi-Supervised Learning, Open-set Semi-Supervised Learning

TL;DR: This paper introduces Federated Open-Set Semi-Supervised Learning (FOSSL) under labels-at-server, identifies its challenges, and proposes a global-pivot–centric framework that achieves stable gains in ID accuracy and OOD detection.

Abstract:

We introduce Federated Open-Set Semi-Supervised Learning (FOSSL), a new and practically important federated learning setting where the server holds a small labeled set of in-distribution (ID) classes while clients provide only unlabeled, non-IID data that may include unknown classes. This setting is under-explored and presents two key challenges: pseudo-label brittleness under distributed OOD contamination and amplified heterogeneity arising from diverse OOD categories across clients. These challenges cause conventional federated SSL or centralized OSSL pipelines to become unstable when applied directly.
We propose OpenFL, a server-guided framework designed to remain robust under these FOSSL-specific difficulties. OpenFL stabilizes global training via a round-wise EMA model, maintains class-level pivots as global anchors for representation learning, and aggregates clients using reliability-aware weights. Clients perform gated pivot alignment, strengthening ID-consistent updates while suppressing the influence of uncertain or OOD-prone samples.
Across CIFAR-10, CIFAR-100, and FashionMNIST with diverse inlier/outlier splits and unseen OOD tests, OpenFL improves both ID accuracy and OOD detection while maintaining stable training. This work establishes FOSSL as a benchmark problem and provides a principled framework for learning under unlabeled, open-set, and highly heterogeneous federated environments.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 18227