ProxyFL: A Proxy-Guided Framework for Federated Semi-Supervised Learning
Duowen Chen, Yan Wang
TL;DR
ProxyFL introduces a novel proxy-guided framework for federated semi-supervised learning that effectively mitigates data heterogeneity across and within clients, leading to improved model performance and convergence.
Contribution
It proposes a unified proxy-based approach to address both external and internal heterogeneity in FSSL, enhancing data utilization and model accuracy.
Findings
Significant performance improvements over existing methods.
Effective mitigation of data heterogeneity issues.
Proven convergence through theoretical analysis.
Abstract
Federated Semi-Supervised Learning (FSSL) aims to collaboratively train a global model across clients by leveraging partially-annotated local data in a privacy-preserving manner. In FSSL, data heterogeneity is a challenging issue, which exists both across clients and within clients. External heterogeneity refers to the data distribution discrepancy across different clients, while internal heterogeneity represents the mismatch between labeled and unlabeled data within clients. Most FSSL methods typically design fixed or dynamic parameter aggregation strategies to collect client knowledge on the server (external) and / or filter out low-confidence unlabeled samples to reduce mistakes in local client (internal). But, the former is hard to precisely fit the ideal global distribution via direct weights, and the latter results in fewer data participation into FL training. To this end, we…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1.The paper has a strong focus on addressing key pain points: it centers on the core bottlenecks of external heterogeneity and internal heterogeneity in FSSL and proposes corresponding solutions. The comparison methods encompass three categories: FL, FL+SSL and FSSL, ensuring comparability. 2.This paper introduces the concept of “classifier weights as a unified proxy”, simultaneously addressing both internal and external heterogeneity. The GPT optimization logic is reasonably sound, the ICPL des
1.The core differences between ProxyFL and PCL have not been sufficiently elaborated. It is necessary to explain the scenario adaptation innovations achieved by applying PCL to FSSL, rather than merely a direct combination of methods. 2.The advantage of "unification" has not been verified.The performance difference between the "unified proxy” and the hybrid schemes of “GPT + existing internal heterogeneity processing method” or “existing external heterogeneity processing method + ICPL” has not b
1. Unified treatment of heterogeneity. The proxy formulation provides a single mechanism to address both cross‑client (external) and within‑client (internal) distributional challenges. 2. Role separation between server and clients. The server focuses on global proxy tuning, while clients use indecisive‑category proxy learning—a clear division of responsibilities. 3. Empirical gains. The method demonstrates competitive or superior results in the presented experiments, with indications of faster
1. **Positioning of global proxy tuning.** The paper should more clearly articulate how the proposed global proxy differs from prior server‑side prototype/centroid refinement—e.g., Orchestra’s globally consistent clustering and centroid updates [1]. - What is unique about optimizing classifier weight vectors as proxies (objective, constraints, and update rules) versus updating global centroids/prototypes? - Are there theoretical or empirical reasons why class‑weight proxies better capture global
1.Originality: The use of proxies to address both external and internal heterogeneity in FSSL is a highly original and valuable contribution. The idea of using learnable classifier weights as proxies to model category distributions both locally and globally is innovative and offers a scalable solution. 2.Quality: The experiments are thorough and demonstrate significant improvements over existing methods. The paper presents a clear explanation of the ProxyFL framework, backed by extensive empiric
1. While the proposed ProxyFL framework shows great potential, the theoretical discussion could be further expanded. Specifically, a more detailed explanation of why proxies are effective in mitigating both external and internal heterogeneity would strengthen the paper. It would be helpful to explore how proxies help the model adapt to data distributions under varying levels of heterogeneity. 2.While the paper introduces key mechanisms like the Dynamic Proxy Pool and Global Proxy Tuning , these
- Addresses both external and internal heterogeneity with a unified proxy. - Paper is well written and easy to follow.
- Requiring knowledge of global labeled-data distribution may still raise privacy concerns, particularly for rare class distributions or small client populations. - Efficiency evaluation is missing. ProxyFL introduces additional server-side computation and utilizing more unlabeled data, which results in more computation. Communication overhead and convergence analysis with wall-clock should be compared against baselines, not only epoch-level metrics. - Lacks comparison with existing method. For
1. The authors proposed two techniques for Federated semi-supervised learning to mitigate heterogeneity, namely a Global Proxy-Tuning (GPT) and IndecisiveCategories Proxy Learning (ICPL) . The topic is relevant and important 2. The presentation is clear. The author generally have a clean problem formulation and math notation. 3. Clean thought flow. The author starts from observations from experiments, then raised questions and propose solution. 4. Experimental results and abliation study are s
1. I respectfully believe, the contribution of the methods is slightly low. For Global Proxy-Tuning (GPT), the idea -- averaging local weights with the considerations of outliers--, have been expored by multiple existing works for federated supervised learning. The Indecisive Categories Proxy Learning (ICPL) is essentially a coarse labeling process, commonly adopted in semi-supervised learning. 2. I would ecourage author briefly analyze the impact of two methods on global learning dynamic, assu
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPrivacy-Preserving Technologies in Data · Domain Adaptation and Few-Shot Learning · Machine Learning and Data Classification