FedUDG: Federated Domain Generalization with Unlabeled Clients

Junyu Ren; Wensheng Gan; Philip S. Yu

cs.LGNov 1, 2026

FedUDG: Federated Domain Generalization with Unlabeled Clients

Authors: Junyu Ren, Wensheng Gan, Philip S. Yu

Organizations: College of Cyber Security, Jinan University, Guangzhou 510632, China. · Department of Computer Science, University of Illinois Chicago, Chicago, USA.

Abstract

Federated domain generalization (FDG) aims to learn models that generalize reliably to out-of-distribution (OOD) environments without centralizing client's data. However, existing FDG methods assume that all source clients have labeled data, limiting their ability to leverage the abundant unlabeled client data available in practice. To address this limitation, we propose FedUDG. At the federation level, FedUDG utilizes a parameter-functional differentiation strategy that maintains a semantic anchor updated solely using ground-truth labels. The remaining parameters absorb knowledge from unlabeled domains under bounded aggregation, preventing domain-biased noise from eroding the class semantics. Within each unlabeled client, FedUDG utilizes a student-teacher framework with two complementary objectives: (1) semantic self-training on high-confidence samples, using the functional differentiation between the semantic anchor and classifier to constrain the model to learn semantically pure and cross-domain stable features; and (2) label-free representation distillation on all unlabeled samples to stabilize learning under appearance variations. Together, these objectives guide the model toward a transferable feature space, while mitigating overfitting to domain-specific characteristics. Furthermore, to address the lack of systematic evaluation of FDG under formal privacy constraints, we develop a sample-level differentially private extension, DP-FedUDG, to quantify the trade-off between privacy and OOD generalization. Across four evaluation protocols, FedUDG consistently outperforms state-of-the-art baselines. The code is available at https://github.com/Jr10426/FedUDG

Figures & tables

Fig. 1: Comparison of the problem settings of FSSL, FDG, and this work. FSSL jointly exploits labeled and unlabeled data but typically evaluates on the client distributions seen during training. FDG targets OOD generalization to unseen target clients but generally assumes that all source clients are labeled. FedUDG addresses both heterogeneous label availability across source clients and generalization to an unseen OOD target client.

Fig. 2: Overall architecture of FedUDG. Labeled clients learn the encoder, classifier, and semantic anchors from ground-truth labels. Within a student-teacher framework, unlabeled clients perform semantic self-training on credible samples and label-free representation distillation on all samples. The server follows the parameter-functional differentiation strategy when aggregating model components and updates the EMA teacher.

Parameter	Value
Backbone	ImageNet-pretrained ResNet-18
Input resolution	$224\times 224$
Communication rounds	80
Local epochs	5
Batch size	16
Optimizer	SGD

TABLE I: Model and Training Configuration

	$M=(3,6)$					$M=(6,12)$					$M=(12,24)$					$M=(12,24),\ \alpha=0.5$
Method	P	A	C	S	Mean	P	A	C	S	Mean	P	A	C	S	Mean	P	A	C	S	Mean
FedIRM [ 12 ]	93.41	73.14	72.78	74.21	78.39	92.57	73.71	72.01	73.56	77.96	90.16	72.68	71.17	72.59	76.65	87.51	71.93	68.12	62.24	72.45
FedPDM [ 35 ]	91.02	76.46	77.60	79.15	81.06	87.66	73.49	75.73	64.95	75.46	86.11	64.75	72.10	45.74	67.18	81.98	58.15	68.26	39.35	61.94
FedRVR [ 34 ]	95.69	81.15	74.57	65.69	79.28	93.89	79.98	74.70	59.00	76.89	93.15	76.98	70.59	56.21	74.23	92.34	74.35	65.23	45.29	69.30
RSCFed [ 14 ]	91.44	75.39	78.11	72.00	79.24	92.93	77.25	75.64	70.88	79.18	93.83	75.44	75.17	61.64	76.52	93.89	69.68	70.56	60.75	73.72
FedGA [ 10 ]	93.86	81.27	76.84	82.83	83.70	91.41	79.82	75.18	83.23	82.41	91.08	77.27	73.97	80.45	80.69	90.95	74.23	71.62	79.90	79.18

TABLE II: Classification Accuracy (%) on Held-Out PACS Targets in the PACS–DomainNet Joint Dataset.

M=(M_{L},M_{U})

denotes the labeled and unlabeled client counts. Best and second-best results are bold and underlined, respectively.

	$\alpha=10$							$\alpha=1$							$\alpha=0.1$
Method	S	R	Q	C	P	I	Mean	S	R	Q	C	P	I	Mean	S	R	Q	C	P	I	Mean
FedIRM [ 12 ]	76.54	81.97	33.25	86.30	85.18	43.35	67.77	76.01	81.52	32.75	85.67	85.03	41.36	67.06	70.19	72.21	30.25	68.15	80.42	34.20	59.24
FedPDM [ 35 ]	78.72	83.78	33.86	88.15	83.14	45.39	68.84	77.65	84.51	33.26	87.58	84.09	44.12	68.54	69.74	74.38	31.98	70.36	82.29	38.53	61.21
FedRVR [ 34 ]	80.97	93.13	37.09	91.72	86.95	46.45	72.72	79.57	92.82	35.99	87.23	85.82	44.15	70.93	62.36	88.95	38.21	69.20	79.41	47.09	64.20
RSCFed [ 14 ]	81.19	89.22	34.57	90.61	85.17	45.11	70.98	80.90	87.50	36.51	89.10	87.00	44.55	70.93	61.50	55.10	39.00	70.90	76.80	48.20	58.58
FedGA [ 10 ]	84.01	94.08	56.29	92.61	88.62	43.98	76.60	82.70	93.95	54.71	89.33	87.91	43.25	75.31	64.26	87.19	49.04	73.58	85.92	42.59	67.10

TABLE III: Classification Accuracy (%) on miniDomainNet under Within-Domain Dirichlet Partitions with

(M_{L},M_{U})=(5,10)

. Best and second-best results are bold and underlined, respectively.

Fig. 3: t-SNE visualization of encoder features for the Photo target in the PACS–DomainNet joint dataset ( $M=(6,12)$ ).

	$\alpha=10$					$\alpha=1$					$\alpha=0.1$
Method	V	L	C	S	Mean	V	L	C	S	Mean	V	L	C	S	Mean
FedIRM [ 12 ]	62.10	51.57	91.81	67.92	68.35	47.83	53.52	90.16	66.58	64.52	36.49	50.17	87.32	51.21	56.30
FedRVR [ 34 ]	62.34	51.26	97.17	58.86	67.41	61.63	50.98	95.71	55.85	66.04	31.29	48.42	94.54	44.67	54.73
FedPDM [ 35 ]	65.72	53.09	93.44	68.57	70.21	63.15	49.11	93.24	66.19	67.92	35.32	48.27	83.30	49.16	54.01
RSCFed [ 14 ]	66.32	59.45	95.95	68.72	72.61	58.90	60.34	96.15	68.06	70.86	35.22	55.47	92.09	62.67	61.36
FedGA [ 10 ]	68.81	55.89	94.35	70.70	72.44	65.82	54.98	93.45	68.67	70.73	35.62	49.63	89.14	52.48	56.72

TABLE IV: Classification Accuracy (%) on VLCS under Within-Domain Dirichlet Partitions with

(M_{L},M_{U})=(3,6)

. Best and second-best results are bold and underlined, respectively.

	PACS					miniDomainNet							VLCS
Method	P-S	A-P	C-A	S-C	Mean	R-S	P-R	C-Q	S-C	I-P	Q-I	Mean	V-S	L-V	C-L	S-C	Mean
FedIRM [ 12 ]	28.81	86.96	63.81	49.43	57.25	78.78	89.82	55.05	71.61	64.52	17.60	62.90	68.12	56.36	46.39	59.48	57.59
FedPDM [ 35 ]	27.38	94.67	74.95	44.11	60.28	79.93	93.97	33.65	87.55	89.09	23.39	67.93	74.84	60.81	37.17	60.54	58.34
FedRVR [ 34 ]	41.74	92.93	66.30	47.52	62.12	81.64	91.86	47.57	83.12	43.76	15.13	60.51	71.83	52.26	23.37	76.59	56.01
RSCFed [ 14 ]	29.19	95.80	64.74	57.59	61.83	79.81	95.47	42.37	80.24	71.95	20.41	65.04	74.27	61.63	57.67	61.55	63.78
FedUDG	35.99	96.04	76.61	67.02	68.92	82.13	94.19	62.38	85.09	85.19	23.85	72.14	74.01	63.44	52.56	70.64	65.16

TABLE V: Classification Accuracy (%) with One Labeled Source Client on PACS, miniDomainNet, and VLCS. Best and second-best results are bold and underlined, respectively.

Variant	P	A	C	S	Mean
FedAvg	92.56	77.08	77.87	81.11	82.15
FedUDG w/o unlabeled	92.29	80.84	76.56	82.81	83.13
FedUDG w/o $\mathcal{L}_{\mathrm{sem}}$	94.01	83.01	78.97	84.58	85.14
FedUDG w/o $\mathcal{L}_{\mathrm{Dis}}$	95.51	85.24	77.99	84.86	85.90
FedUDG w/o JS	96.07	84.68	77.01	85.28	85.76
FedUDG w/o anchor & JS	95.43	82.56	77.66	84.87	85.13

TABLE VI: Ablation Results (%) on Held-Out PACS Targets in the PACS–DomainNet Joint Dataset with

M=(3,6)

Fig. 4: Pseudo-label precision among admitted samples (a) and admission coverage (b) on the PACS–miniDomainNet joint dataset. Lines show the mean across four target tasks; shaded bands show the minimum–maximum range.

Parameter	Value	P	A	C	S	Mean
$\tau$	0.7	95.51	85.01	79.18	85.03	86.18
	0.8	95.63	85.69	78.75	86.21	86.57
	0.9	94.97	84.13	78.71	85.42	85.81
$\kappa$	0.05	94.85	85.74	78.71	85.14	86.11
	0.1	95.63	85.69	78.75	86.21	86.57
	0.2	94.92	84.90	78.09	85.03	85.74

TABLE VII: Parameter Sensitivity (%) on Held-Out PACS Targets in the PACS–DomainNet Joint Dataset with

M=(3,6)

Method	$\sigma$	$\varepsilon_{\max}$	P	A	C	S	Mean
FedUDG	–	$\infty$	95.63	85.69	78.75	86.21	86.57
DP-FedUDG	1.0	16.94	73.59	53.11	42.92	57.25	56.72
	1.5	8.46	53.95	39.51	34.02	41.72	42.30
	2.0	5.66	29.46	20.34	22.99	25.35	24.54

TABLE VIII: Accuracy (%) on Held-Out PACS Targets under Sample-Level DP in the PACS–DomainNet Joint Dataset (

\delta=10^{-5}

C=1

Federated Domain Generalization

See all 13 papers

Oct 14, 2025

FedHUG: Federated Heterogeneous Unsupervised Generalization for Remote Physiological Measurements

Remote physiological measurement gained wide attention, while it requires collecting users' privacy-sensitive information, and existing contactless measurements still rely on labeled client data. This presents challenges when we want to further update real-world deployed models with numerous user data lacking labels. To resolve these challenges, we instantiate a new protocol called Federated Unsupervised Domain Generalization (FUDG) in this work. Subsequently, the Federated Heterogeneous Unsupervised Generalization (FedHUG) framework is proposed and consists of: (1) Minimal Bias Aggregation module dynamically adjusts aggregation weights based on prior-driven bias evaluation to cope with heterogeneous non-IID features from multiple domains. (2) The Global Distribution-aware Learning Controller parameterizes the label distribution and dynamically manipulates client-specific training strategies, thereby mitigating the server-client label distribution skew and long-tail issue. The proposal shows superior performance across state-of-the-art techniques in estimation with either RGB video or mmWave radar. The code will be released.

Xiao Yang, Dengbo He, Duoxun Tang +2

Systems Hub, the Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Federated Domain Generalization Federated Learning Federated Learning Aggregation Heterogeneous Federated Learning Non-IID Federated Learning Remote Photoplethysmography

Ask PDF

Sep 15, 2025

FedDAF: Federated Domain Adaptation Using Model Functional Distance

Federated Domain Adaptation (FDA) improves model performance at a target client by collaborating with source clients while preserving data privacy. FDA faces two key challenges: domain shift between source and target data, and limited labeled data at the target, a common constraint when a new site joins a federation before it has accumulated its own labeled data, as in clinical deployments. Most existing methods address domain shift alone, assuming ample target data; those that also tackle data scarcity still fail to prioritize source information according to the target's specific objective. We propose FedDAF, which addresses both challenges through similarity-based aggregation of the global source and target models, using their model functional distance, computed from the angle between their mean gradient fields on target data and normalized via a Gompertz function. The global source model itself is formed using a distance-based weighted average, giving greater weight to source models closer to the target model. Experiments on real-world datasets show FedDAF outperforms existing federated learning (FL), personalized FL, and FDA methods in test accuracy.

Mrinmay Sen, Sidhant Nair, C Krishna Mohan

Department of Artificial Intelligence, Indian Institute of Technology Hyderabad, Hyderabad, 502285, Telangana, India. · Department of Mechanical Engineering, Indian Institute of Technology Delhi, New Delhi, 110016, Delhi, India.

Federated Learning Aggregation Domain Adaptation Model Merging Federated Learning

Ask PDF

Mar 15, 2026

Domain-Skewed Federated Learning with Feature Decoupling and Calibration

Federated learning (FL) allows distributed clients to collaboratively train a global model in a privacy-preserving manner. However, one major challenge is domain skew, where clients' data originating from diverse domains may hinder the aggregated global model from learning a consistent representation space, resulting in poor generalizable ability in multiple domains. In this paper, we argue that the domain skew is reflected in the domain-specific biased features of each client, causing the local model's representations to collapse into a narrow low-dimensional subspace. We then propose Federated Feature Decoupling and Calibration (

F^2

DC), which liberates valuable class-relevant information by calibrating the domain-specific biased features, enabling more consistent representations across domains. A novel component, Domain Feature Decoupler (DFD), is first introduced in

F^2

DC to determine the robustness of each feature unit, thereby separating the local features into domain-robust features and domain-related features. A Domain Feature Corrector (DFC) is further proposed to calibrate these domain-related features by explicitly linking discriminative signals, capturing additional class-relevant clues that complement the domain-robust features. Finally, a domain-aware aggregation of the local models is performed to promote consensus among clients. Empirical results on three popular multi-domain datasets demonstrate the effectiveness of the proposed

F^2

DC and the contributions of its two modules. Code is available at https://github.com/mala-lab/F2DC.

Huan Wang, Jun Shen, Jun Yan +1

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

Federated Learning Heterogeneous Federated Learning Domain Generalization Representation Disentanglement

Ask PDF

Abstract

Federated Domain Generalization

Explore similar work

FedHUG: Federated Heterogeneous Unsupervised Generalization for Remote Physiological Measurements

FedDAF: Federated Domain Adaptation Using Model Functional Distance

Domain-Skewed Federated Learning with Feature Decoupling and Calibration