OpenNovelty.org

FedOpenMatch: Towards Semi-Supervised Federated Learning in Open-Set Environments

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 5.5Download Report PDF
semi-supervised federated learningopen-setfederated learning
Abstract:

Semi-supervised federated learning (SSFL) has emerged as an effective approach to leverage unlabeled data distributed across multiple data owners for improving model generalization. Existing SSFL methods typically assume that labeled and unlabeled data share the same label space. However, in realistic federated scenarios, unlabeled data often contain categories absent from the labeled set, i.e., outliers, which can severely degrade the performance of SSFL algorithms. In this paper, we address this under-explored issue, formally propose the open-set semi-supervised federated learning (OSSFL) problem, and develop the first OSSFL framework, FedOpenMatch. Our method adopts a one-vs-all (OVA) classifier as the outlier detector, equipped with logit adjustment to mitigate inlier-outlier imbalance and a gradient stop mechanism to reduce feature interference between the OVA and inlier classifiers. In addition, we introduce the logit consistency regularization loss, yielding more robust performance. Extensive experiments on standard benchmarks across diverse data settings demonstrate the effectiveness of FedOpenMatch, which significantly outperforms the baselines.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces FedOpenMatch to address open-set semi-supervised federated learning (OSSFL), where unlabeled data contain novel classes absent from labeled sets. According to the taxonomy, this work resides in the 'Open-World Semi-Supervised Federated Learning' leaf, which contains only two papers total. This represents a sparse research direction within the broader field of 21 surveyed papers across 10 leaf nodes, suggesting the problem space is relatively under-explored compared to neighboring areas like federated generalized category discovery (4 papers) or federated novel class discovery (2 papers).

The taxonomy reveals that neighboring research directions address related but distinct challenges. Federated generalized category discovery focuses on clustering unlabeled data into known and novel classes, while federated novel class discovery emphasizes incorporating new categories during training. The paper's positioning in open-world semi-supervised learning distinguishes it through explicit handling of outlier detection in unlabeled data, as indicated by the leaf's scope note emphasizing 'debiasing or unbiased training strategies.' This contrasts with category discovery methods that assume explicit clustering objectives rather than outlier rejection.

Among the three contributions analyzed, the OSSFL problem formulation examined 9 candidates with 1 appearing to provide overlapping prior work, while the FedOpenMatch framework examined only 1 candidate with no clear refutation. The baseline implementations contribution examined 10 candidates with 1 refutable match. Given the limited search scope of 20 total candidates examined, these statistics suggest the framework components appear relatively novel, though the problem formulation and baseline adaptations have some precedent in the examined literature. The single sibling paper in the same taxonomy leaf likely represents the most directly comparable prior work.

Based on the top-20 semantic matches examined, the work appears to occupy a genuinely sparse research area, with limited direct competition in the specific OSSFL formulation. However, the analysis cannot assess whether related work exists outside this search scope, particularly in adjacent areas like open-set domain adaptation or continual federated learning that might address similar challenges through different problem framings.

Taxonomy

Core-task Taxonomy Papers
21
3
Claimed Contributions
20
Contribution Candidate Papers Compared
2
Refutable Paper

Research Landscape Overview

Core task: open-set semi-supervised federated learning. This emerging field addresses scenarios where distributed clients collaboratively train models on partially labeled data while encountering novel classes not present in the labeled set. The taxonomy reveals a rich structure spanning five main branches. Core Open-Set Semi-Supervised Federated Learning Methods tackle the fundamental challenge of discovering and handling unknown categories in federated settings, with works like FedOpenMatch[0] and Unbiased Federated Open-World[6] directly addressing open-world recognition under label scarcity. Specialized Federated Learning Variants explore domain-specific adaptations such as graph-based learning (GraphFL[1]) and positive-unlabeled scenarios (Federated Positive Unlabeled[2]). Continual and Lifelong Federated Learning focuses on sequential learning with evolving class sets, exemplified by Long-term Federated Continual[10] and Federated Continuous Discovery[11]. Cross-Domain and Out-of-Distribution Detection emphasizes robustness to distribution shifts (OOD Detection Wild[18]), while Supporting Techniques and Applications provide practical implementations across diverse domains from audio (Zero-shot Federated Audio[4]) to medical imaging (SODA COVID Detection[13]). A particularly active line of work centers on balancing the dual challenges of semi-supervised learning and novel class discovery in federated environments. Methods like Distillation Open Semi-supervised[3] and Federated Category Discovery[5] explore knowledge transfer and clustering techniques to identify emerging categories, while FedOpenMatch[0] sits at the intersection of these themes by proposing mechanisms to match unlabeled samples across clients while detecting out-of-distribution instances. Compared to Unbiased Federated Open-World[6], which emphasizes debiasing strategies for open-world recognition, FedOpenMatch[0] appears to focus more directly on the semi-supervised matching problem under open-set conditions. Trade-offs between communication efficiency, privacy preservation, and discovery accuracy remain central open questions, with recent works like Federated Novel Prompts[14] and Cloud-edge Category Discovery[15] exploring prompt-based and hierarchical architectures to address scalability challenges in real-world deployments.

Claimed Contributions

Open-set semi-supervised federated learning (OSSFL) problem formulation

The authors formally define the OSSFL problem, which addresses the realistic scenario where unlabeled data distributed across federated clients contain samples from classes absent in the server's labeled dataset. This extends open-set semi-supervised learning to federated settings.

9 retrieved papers
Can Refute
FedOpenMatch framework with gradient stop, logit adjustment, and logit consistency regularization

The authors propose FedOpenMatch, a multi-task learning framework that jointly trains an inlier classifier and an OVA-based outlier detector. The framework incorporates gradient stop to decouple feature spaces, logit adjustment to handle inlier-outlier imbalance, and logit consistency regularization to exploit unlabeled samples more effectively.

1 retrieved paper
OSSFL baseline implementations by adapting OSSL algorithms

The authors create federated counterparts of recent centralized open-set SSL methods (OpenMatch, SSB, IOMatch, BDMatch) to enable fair comparison in the OSSFL setting, since no prior OSSFL methods exist.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Open-set semi-supervised federated learning (OSSFL) problem formulation

The authors formally define the OSSFL problem, which addresses the realistic scenario where unlabeled data distributed across federated clients contain samples from classes absent in the server's labeled dataset. This extends open-set semi-supervised learning to federated settings.

Contribution

FedOpenMatch framework with gradient stop, logit adjustment, and logit consistency regularization

The authors propose FedOpenMatch, a multi-task learning framework that jointly trains an inlier classifier and an OVA-based outlier detector. The framework incorporates gradient stop to decouple feature spaces, logit adjustment to handle inlier-outlier imbalance, and logit consistency regularization to exploit unlabeled samples more effectively.

Contribution

OSSFL baseline implementations by adapting OSSL algorithms

The authors create federated counterparts of recent centralized open-set SSL methods (OpenMatch, SSB, IOMatch, BDMatch) to enable fair comparison in the OSSFL setting, since no prior OSSFL methods exist.

Table of Contents