FedDBP: Enhancing Federated Prototype Learning with Dual-Branch Features and Personalized Global Fusion

Ningzhi Gao; Siquan Huang; Leyu Shi; Ying Gao

cs.CVMar 31, 2026

FedDBP: Enhancing Federated Prototype Learning with Dual-Branch Features and Personalized Global Fusion

Authors: Ningzhi Gao, Siquan Huang, Leyu Shi, Ying Gao

Organizations: South China University of Technology, Guangzhou, China

Abstract

Federated prototype learning (FPL), as a solution to heterogeneous federated learning (HFL), effectively alleviates the challenges of data and model heterogeneity.However, existing FPL methods fail to balance the fidelity and discriminability of the feature, and are limited by a single global prototype. In this paper, we propose FedDBP, a novel FPL method to address the above issues. On the client-side, we design a Dual-Branch feature projector that employs L2 alignment and contrastive learning simultaneously, thereby ensuring both the fidelity and discriminability of local features. On the server-side, we introduce a Personalized global prototype fusion approach that leverages Fisher information to identify the important channels of local prototypes. Extensive experiments demonstrate the superiority of FedDBP over ten existing advanced methods.

Figures & tables

	CIFAR-10		CIFAR-100		Flowers102		Tiny-ImageNet
FedProto [AAAI22]	$\small 84.32\pm\pm 0.24$		$\small 39.01\pm\pm 0.17$		$\small 43.61\pm\pm 0.37$		$\small 19.33\pm\pm 0.08$
FedProc [FCGS23]	$\small 86.14\pm\pm 0.19$	$\small\uparrow$ 1.82	$\small 41.12\pm\pm 0.26$	$\small\uparrow$ 2.11	$\small 44.00\pm\pm 0.53$	$\small\uparrow$ 0.39	$\small 23.54\pm\pm 0.35$	$\small\uparrow$ 4.21
FPL [CVPR23]	$\small 86.08\pm\pm 0.14$	$\small\uparrow$ 1.76	$\small 40.15\pm\pm 0.12$	$\small\uparrow$ 1.04	$\small 40.33\pm\pm 0.20$	$\small\downarrow$ 3.28	$\small 20.87\pm\pm 0.15$	$\small\uparrow$ 1.54
FedTGP [AAAI24]	$\small 88.04\pm\pm 0.21$	$\small\uparrow$ 3.72	$\small 41.77\pm\pm 0.27$	$\small\uparrow$ 2.76	$\small 51.74\pm\pm 0.31$	$\small\uparrow$ 8.13	$\small 24.72\pm\pm 0.11$	$\small\uparrow$ 5.39
FedLFP [TMC25]	$\small 85.76\pm\pm 0.09$	$\small\uparrow$ 1.44	$\small 39.80\pm\pm 0.18$	$\small\uparrow$ 0.79	$\small 47.38\pm\pm 0.23$	$\small\uparrow$ 3.77	$\small 23.37\pm\pm 0.21$	$\small\uparrow$ 4.04
FD [arXiv18]	$\small 86.61\pm\pm 0.11$	$\small\uparrow$ 2.29	$\small 41.42\pm\pm 0.14$	$\small\uparrow$ 2.41	$\small 48.69\pm\pm 0.38$	$\small\uparrow$ 5.08	$\small 25.57\pm\pm 0.28$	$\small\uparrow$ 6.24

Table 1: Comparison with state-of-the-art federated prototype learning and other HFL methods on four datasets. We show the average test accuracy(%) across clients for each method. Best in bold and second with underline. The growth or reduction is compared to FedProto.

Figure 2: Visualization of training curves of the average accuracy(%) of FedDBP and other HFL methods on the CIFAR-10 dataset.

Figure 3: Visualization of accuracy(%) of our method and others on different degrees of non-IID, which is indicated by the

\small\alpha

parameter of the Dirichlet distribution.

	E=1	E=10	E=20
FedProto	$\small 34.67$	$\small 39.01$	$\small 38.65$
FedProc	$\small 41.22$	$\small 41.12$	$\small 41.83$
FPL	$\small 38.83$	$\small 40.15$	$\small 39.53$
FedTGP	$\small 38.21$	$\small 41.77$	$\small 43.03$
FedDBP	$\small 42.59$	$\small 45.59$	$\small 45.26$

Table 2: Comparison with federated prototype learning methods on CIFAR-100 dataset with different number of training epochs( E ).

	CIFAR10	CIFAR100	Flowers102	Tiny-ImageNet
w/o share	$\small 87.68$	$\small 40.65$	$\small 42.56$	$\small 19.11$
w/o deci.	$\small 85.95$	$\small 38.42$	$\small 44.3$	$\small 20.81$
w/o hard	$\small 86.94$	$\small 42.76$	$\small 48.16$	$\small 23.22$
w/o pers.	$\small 88.02$	$\small 44.68$	$\small 52.6$	$\small 26.94$
FedDBP	$\small 88.27$	$\small 45.59$	$\small 52.99$	$\small 27.14$

Table 3: Effectiveness of key components on four datasets.

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
FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift
Apr 8, 2026Huy Q. Le, Loc X. Nguyen, Yu Qiao +3Federated Learning Distribution Shift
G-LAMP NEXUS Institute, Kyung Hee University · Kyung Hee University
Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation
Jul 23, 2024Xinghao Wu, Xuefeng Liu, Jianwei Niu +4Prompt Learning Federated Learning
School of Computer Science and Engineering, Beihang University, Beijing, China · Zhongguancun Laboratory, Beijing, China · Sichuan University, Sichuan, China +1

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

FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift

Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation