ProxyFL: A Proxy-Guided Framework for Federated Semi-Supervised Learning

Duowen Chen; Yan Wang

cs.LGFeb 24, 2026

ProxyFL: A Proxy-Guided Framework for Federated Semi-Supervised Learning

Authors: Duowen Chen, Yan Wang

Organizations: Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University

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 propose a proxy-guided framework called ProxyFL that focuses on simultaneously mitigating external and internal heterogeneity via a unified proxy. I.e., we consider the learnable weights of classifier as proxy to simulate the category distribution both locally and globally. For external, we explicitly optimize global proxy against outliers instead of direct weights; for internal, we re-include the discarded samples into training by a positive-negative proxy pool to mitigate the impact of potentially-incorrect pseudo-labels. Insight experiments & theoretical analysis show our significant performance and convergence in FSSL.

Figures & tables

Figure 1: (a) Illustration of centralized learning and averaging-based federated learning. (b) Differences of test accuracy and pseudo-labeling accuracy under varying levels of heterogeneity (smaller $\alpha$ indicates greater heterogeneity). During each communication round, all clients are trained based on FedSGD [ 23 ] w/ , w/o low-confidence samples and our method for one local epoch.

Figure 2: (a) t-SNE visualization of classifier weights from clients (circle) during the initial round. Different colors means different categories. Pentagram denotes the simple average of weight parameters by category. (b) Number of excluded unlabeled samples under different levels of heterogeneity. (c) Test accuracy curves of ‘SGD-FSSL’ w/o and w/ low-confidence samples under different ratios of correct labels. These samples could improve performance as the correctly-labeled number increases.

Figure 3: Overview of ProxyFL for Global Proxy Tuning in server-side & Indecisive-Categories Proxy Learning in local client.

Figure 4: (a-b) Examples of the indecisive-categories. (c) Illustration of our Indecisive-Categories Proxy Learning.

Methods	CIFAR-10			CIFAR-100			SVHN			CINIC-10
Methods	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$
FL Methods
FedAvg [ 23 ]	69.60	68.88	69.39	34.08	33.21	35.31	82.40	83.40	78.60	57.17	60.09	61.54
FedProx [ 17 ]	68.58	69.53	68.00	34.20	34.07	34.88	81.67	83.77	83.77	58.05	60.71	62.82
FedAvg-SL	90.46	91.24	91.32	67.98	68.83	69.10	94.11	94.41	94.40	77.82	80.42	81.29
FL+SSL Methods

Table 1: Experimental results on CIFAR-10, CIFAR-100, SVHN and CINIC-10 under 10% label. Bold text indicates the best result, and the last row presents the improvement of ProxyFL over the second best method.

GPT	ICPL	CIFAR10			CIFAR100			SVHN			CINIC10
GPT	ICPL	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$
		84.56	86.05	86.66	48.96	51.80	52.19	90.50	91.94	92.31	71.67	73.26	74.80
✓		87.59	89.23	89.71	54.86	56.58	57.09	94.29	94.49	94.53	77.15	79.03	79.31
	✓	87.81	89.58	89.66	57.21	57.98	57.74	94.82	94.69	95.15	77.80	78.04	78.57
✓	✓	88.56	90.00	89.96	57.50	58.75	58.24	95.09	95.18	95.26	77.98	78.96	79.59

Table 2: Module ablation studies on GPT and ICPL of our method.

Figure 5: (a) Distribution of labeled and unlabeled data across clients under $\alpha=0.1$ taking CIFAR-10 as an example. (b) Convergence curves of ProxyFL and other baselines on CIFAR-100 with $\alpha=0.1$ . (c) Distribution of global category proxies before-and after-tuning visualized in a t-SNE plot.

Acc.	30%		40%		50%
Method	Round $\downarrow$	Speedup $\uparrow$	Round $\downarrow$	Speedup $\uparrow$	Round $\downarrow$	Speedup $\uparrow$
LPL	119	$\times$ 1.00	242	$\times$ 1.00	562	$\times$ 1.00
GPL	114	$\times$ 1.04	226	$\times$ 1.07	524	$\times$ 1.07
FedLabel	94	$\times$ 1.27	175	$\times$ 1.38	429	$\times$ 1.31
FedDB	103	$\times$ 1.16	206	$\times$ 1.17	-	-
FedDure	114	$\times$ 1.04	234	$\times$ 1.03	542	$\times$ 1.04

Table 3: Comparison of convergence rates between ProxyFL and other baselines with

\alpha=0.1

on CIFAR-100.

Methods	CIFAR-10	CIFAR-100	SVHN	CINIC-10
FedAvg-SL	90.46	67.98	94.11	77.82
GPL	84.56	48.96	90.50	71.67
GPL-ALL	85.38	50.34	93.31	75.83
LPL	82.98	49.32	89.68	68.02
LPL-ALL	87.18	53.28	83.99	73.39
ICPL-Top1	87.13	55.66	94.56	77.01

Table 4: Ablation of design choices between ICPL and other methods under

\alpha=0.1

Figure 6: Ratios of ‘Correct category in $\xi$ ’ vs. ‘Correct pseudo-label’ for low-confidence samples.

Methods	CIFAR-10	CIFAR-100	SVHN	CINIC-10
PROTO	85.89	49.56	93.48	72.53
PROTO-ALL	86.56	54.69	19.59	73.16
PROTO-Labeled	86.55	48.83	93.60	72.57
ICPL - $\mathcal{P}_{\mathcal{G}}^{\prime}(\mathbf{Y})$	87.81	57.21	94.82	77.80

Table 5: Ablation of our ProxyFL vs. Prototypes with

\alpha=0.1

. ’ALL’ means generating prototypes with low-confidence samples, and ’Labeled’ with only labeled samples.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	CIFAR-10			CIFAR-100			SVHN			CINIC-10
Methods	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$	$\alpha=0.1$	$\alpha=0.5$	$\alpha=1$
FL Methods
FedAvg [ 23 ]	86.37	87.06	87.97	45.72	46.57	47.55	88.37	89.05	89.97	66.24	68.29	69.21
FedProx [ 17 ]	86.78	88.11	88.44	45.96	47.33	47.89	87.99	88.56	91.10	65.53	69.57	69.91
FedAvg-SL	90.46	91.24	91.32	67.98	68.83	69.10	94.11	94.41	94.40	77.82	80.42	81.29
FL+SSL Methods

Appendix

Table 6: Experimental results on CIFAR-10, CIFAR-100, SVHN and CINIC-10 under 20% label. Bold text indicates the best result, and the last row presents the improvement of ProxyFL over the second best method.

	Formulation	Practice
FixMatch+FSSL	$\sum_{m=1}^{M}\|\mathbf{\Theta}\|+\|\mathbf{\Theta}\|$	22.08 M
FedMatch	$\sum_{i=1}^{M}(\|\Delta\sigma_{i}\|+\|\Delta\psi_{i}\|+C)+$	33.12 M
FedMatch	$(\|\Delta\sigma_{\mathcal{G}}\|+\|\Delta\psi_{\mathcal{G}}\|+H\times\|\psi_{h}\|))$	33.12 M
FedLabel	$\sum_{m=1}^{M}\|\mathbf{\Theta}\|+\|\mathbf{\Theta}\|$	22.08 M
FedDure	$\sum_{m=1}^{M}\|\mathbf{\Theta}\|+\|\mathbf{\Theta}\|$	22.08 M
FedDB	$\sum_{i=1}^{M}(\|\mathbf{\Theta}\|+C)+\|\mathbf{\Theta}\|$	22.08 M

Appendix

Table 7: The communication costs per round between different methods in our experimental setting as SAGE [ 21 ] .

\mathbf{\Theta}

represents model parameters.

\sigma

and

\psi

denotes the parameters after sparsely decomposition in FedMatch [ 12 ] , and

H

means the number of helper agents.

C

and

d

means respectively the number of categories and feature dimension.

Acc.	30%		40%		50%
Method	Round $\downarrow$	Speedup $\uparrow$	Round $\downarrow$	Speedup $\uparrow$	Round $\downarrow$	Speedup $\uparrow$
LPL	121	$\times$ 1.00	221	$\times$ 1.00	546	$\times$ 1.00
GPL	113	$\times$ 1.07	210	$\times$ 1.05	419	$\times$ 1.30
FedLabel	83	$\times$ 1.46	160	$\times$ 1.38	366	$\times$ 1.49
FedDB	94	$\times$ 1.29	205	$\times$ 1.08	492	$\times$ 1.11
FedDure	110	$\times$ 1.10	222	$\times$ 1.00	552	$\times$ 0.99

Appendix

Table 8: Comparison of convergence rates between ProxyFL and other baseline methods on CIFAR100 dataset with

\alpha=0.5

(the upper part) and

\alpha=1

(the lower part).

Methods	CIFAR-10	CIFAR-100	SVHN	CINIC-10
FixMatch	82.98	49.32	89.68	68.02
+ProxyFL	88.56	57.50	95.09	77.98
+ProxyFL	$\uparrow$ 5.58	$\uparrow$ 8.18	$\uparrow$ 5.41	$\uparrow$ 9.96
FedLabel	62.85	50.88	89.31	67.64
+ProxyFL	88.55	56.88	94.94	76.35
+ProxyFL	$\uparrow$ 25.70	$\uparrow$ 6.00	$\uparrow$ 5.63	$\uparrow$ 8.71

Appendix

Table 9: Performance gains brought by ProxyFL as a plugin to other baseline methods.

Figure 7: Ablation on the numbers of epoch for Global Proxy Tuning.

Federated Learning

Federated learning trains shared models across data holders without requiring them to pool their raw data.

See all 943 papers

FedHPro: Federated Hyper-Prototype Learning via Gradient Matching
May 13, 2026Huan Wang, Jun Shen, Haoran Li +6Federated Learning Heterogeneous Federated Learning
School of Computing and Information Technology, University of Wollongong, Wollongong, Australia · School of Computing and Information Systems, Singapore Management University, Singapore, Singapore · Monash University, Australia +3
Beyond Non-IID: Learner--Client Distribution Mismatch in Federated Learning
Aug 27, 2026Yiming Xie, Lili Su, Ningfang MiClient Selection in Federated Learning Non-IID Federated Learning
Department of Electrical and Computer Engineering Northeastern University, Boston, MA, USA
Cooperative Multi-Agent Reinforcement Learning for Adaptive Aggregation in Semi-Supervised Federated Learning with non-IID Data
Aug 26, 2026Rene Glitza, Luca Becker, Rainer MartinFederated Learning Federated Learning Aggregation
Institute of Communication Acoustics, Ruhr-Universität Bochum, Bochum, Germany

May 13, 2026J/K move · Enter open · S save

FedHPro: Federated Hyper-Prototype Learning via Gradient Matching

Huan Wang, Jun Shen, Haoran Li, Zhenyu Yang+5

School of Computing and Information Technology, University of Wollongong, Wollongong, Australia · School of Computing and Information Systems, Singapore Management University, Singapore, Singapore · Monash University, Australia+3

Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

Federated Learning (FL) enables collaborative training of distributed clients while protecting privacy. To enhance generalization capability in FL, prototype-based FL is in the spotlight, since shared global prototypes offer semantic anchors for aligning client-specific local prototypes. However, existing methods update global prototypes at the prototype-level via averaging local prototypes or refining global anchors, which often leads to semantic drift across clients and subsequently yields a misaligned global signal. To alleviate this issue, we introduce hyper-prototypes, defined by a set of learnable global class-wise prototypes to preserve underlying semantic knowledge across clients. The hyper-prototypes are optimized via gradient matching to align with class-relevant characteristics distilled directly from clients' real samples, rather than prototype-level descriptors. We further propose FedHPro, a Federated Hyper-Prototype Learning framework, to leverage hyper-prototypes to promote inter-class separability via mutual-contrastive learning with client-specific margin, while encouraging intra-class uniformity through a consistency penalty. Comprehensive experiments under diverse heterogeneous scenarios confirm that 1) hyper-prototypes produce a more semantically consistent global signal, and 2) FedHPro achieves state-of-the-art performance on several benchmark datasets. Code is available at https://github.com/mala-lab/FedHPro.

Abstract

Federated Learning

Explore similar work

FedHPro: Federated Hyper-Prototype Learning via Gradient Matching

Beyond Non-IID: Learner--Client Distribution Mismatch in Federated Learning

Cooperative Multi-Agent Reinforcement Learning for Adaptive Aggregation in Semi-Supervised Federated Learning with non-IID Data