Few-for-Many Personalized Federated Learning

Ping Guo; Tiantian Zhang; Xi Lin; Xiang Li; Zhi-Ri Tang; Qingfu Zhang

cs.AIMar 12, 2026

Few-for-Many Personalized Federated Learning

Authors: Ping Guo, Tiantian Zhang, Xi Lin, Xiang Li, Zhi-Ri Tang, Qingfu Zhang

Organizations: City University of Hong Kong · CityU Shenzhen Research Institute · Hong Kong Metropolitan University · Xi’an Jiaotong University ·

Abstract

Personalized Federated Learning (PFL) aims to train customized models for clients with highly heterogeneous data distributions while preserving data privacy. Existing approaches often rely on heuristics like clustering or model interpolation, which lack principled mechanisms for balancing heterogeneous client objectives. Serving $M$ clients with distinct data distributions is inherently a multi-objective optimization problem, where achieving optimal personalization ideally requires $M$ distinct models on the Pareto front. However, maintaining $M$ separate models poses significant scalability challenges in federated settings with hundreds or thousands of clients. To address this challenge, we reformulate PFL as a few-for-many optimization problem that maintains only $K$ shared server models ( $K \ll M$ ) to collectively serve all $M$ clients. We prove that this framework achieves near-optimal personalization: the approximation error diminishes as $K$ increases and each client's model converges to each client's optimum as data grows. Building on this reformulation, we propose FedFew, a practical algorithm that jointly optimizes the $K$ server models through efficient gradient-based updates. Unlike clustering-based approaches that require manual client partitioning or interpolation-based methods that demand careful hyperparameter tuning, FedFew automatically discovers the optimal model diversity through its optimization process. Experiments across vision, NLP, and real-world medical imaging datasets demonstrate that FedFew, with just 3 models, consistently outperforms other state-of-the-art approaches. Code is available at https://github.com/pgg3/FedFew.

Figures & tables

Figure 1: Paradigms of Personalized Federated Learning. Left: Centralized methods maintain a single global model for all $M$ clients, failing to capture client heterogeneity. Center: Per-client methods train $M$ independent models, sacrificing collaborative learning benefits and suffering from data scarcity. Right: Our proposed few-for-many approach maintains $K$ server models ( $K\ll M$ ) that collectively serve all clients. Each client selects the best-fitting model, achieving strong personalization while preserving collaboration.

Centralized Methods (Single Global Model)
Method	Pathological heterogeneous setting		Practical heterogeneous setting
	CIFAR-100		FEMNIST	CIFAR-10		CIFAR-100		TINY	AG News
	$M=10$	$M=20$	$M=20$	$M=10$	$M=20$	$M=10$	$M=20$	$M=10$	$M=20$
FedAvg [ 29 ]	$\text{29.00}\pm\text{3.94}$	$\text{28.57}\pm\text{4.37}$	$\text{96.65}\pm\text{1.81}$	$\text{61.36}\pm\text{8.54}$	$\text{61.26}\pm\text{8.64}$	$\text{30.84}\pm\text{2.26}$	$\text{31.10}\pm\text{4.21}$	$\text{13.49}\pm\text{1.55}$	$\text{88.90}\pm\text{7.15}$
FedProx [ 22 ]	$\text{28.56}\pm\text{4.50}$	$\text{28.11}\pm\text{3.86}$	$\text{96.51}\pm\text{2.32}$	$\text{60.87}\pm\text{7.87}$	$\text{61.38}\pm\text{9.35}$	$\text{30.74}\pm\text{2.25}$	$\text{30.90}\pm\text{4.03}$	$\text{13.64}\pm\text{1.56}$	$\text{83.42}\pm\text{11.34}$
FedMTL [ 37 ]	$\text{65.33}\pm\text{3.64}$	$\text{59.65}\pm\text{3.70}$	$\textbf{\text{100.00}}\boldsymbol{\pm}\textbf{\text{0.00}}$	$\text{85.92}\pm\text{11.18}$	$\text{85.75}\pm\text{11.62}$	$\text{46.28}\pm\text{3.82}$	$\text{44.79}\pm\text{5.25}$	$\text{23.49}\pm\text{2.43}$	$\text{94.10}\pm\text{7.56}$

Table 1: Overall performance comparison across CIFAR-10, CIFAR-100, TINY (TinyImageNet), AG News, and FEMNIST datasets under pathological and practical heterogeneous settings. Results are reported as mean accuracy (%)

\pm

standard deviation. Best results are highlighted and bolded , second-best results are underlined .

Figure 3: Mean client accuracy comparison across communication configurations. All configurations achieve comparable mean client accuracy (87.8–88.3%), demonstrating that our method is robust to different communication-computation trade-offs.

Figure 4: Communication-computation trade-off. Convergence of

g^{\text{STCH-Set}}

vs total local updates for different (local epochs, communication rounds) configurations. Local epochs (LE)

\in\{1,2,4,8,16\}

with corresponding communication rounds (GR) to maintain 2000 total updates.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description
$M$	number of heterogeneous clients
$K$	number of shared models in the K-for-M framework
$L_{i}(\theta)$	expected (population) loss for client $i$ with model $\theta$
$\hat{L}_{i}(\theta)$	empirical loss for client $i$ based on finite samples
$\theta_{i}^{*}$	optimal personalized model for client $i$ , defined as $\arg\min_{\theta}L_{i}(\theta)$
$\Theta^{(K)}_{*}$	optimal K-for-M solution (minimizing population losses)

Appendix

Table 3: Key notation for theoretical analysis. Stars ( ∗ ) denote optimal or population-level quantities, while hats (

\hat{\cdot}

) denote empirical estimators.

Type	Dataset	Model	Rounds	Local Epochs	Batch Size	Learning Rate	Join Ratio
Benchmark	CIFAR-10	CNN	2000	1	50	0.005	1.0
	CIFAR-100	CNN	2000	1	50	0.005	1.0
	TinyImageNet	CNN	2000	1	50	0.0005	1.0
	AG News	TextCNN	2000	1	100	0.005	1.0
	FEMNIST	CNN	2000	1	100	0.005	1.0
Medical	Kvasir	ResNet-18	1000	1	100	0.002	1.0

Appendix

Table 4: Training hyperparameters for benchmark and medical datasets. Benchmark datasets use 2000 rounds with CNN/TextCNN backbones, while medical datasets use 1000 rounds with ResNet-18 to prevent overfitting on smaller institutional samples. All experiments employ single local epoch and full client participation for fair comparison across methods.

Figure 7: Impact of

\mu

on dual-layer weights. (a) Outer weights

\alpha_{i}

distribution across clients. (b) Inner weight entropy (higher = softer selection). (c) Max inner weight (closer to 1 = harder selection). As

\mu\to 0

, FedFew recovers hard clustering like IFCA.

Figure 8: Convergence vs communication rounds. Re-plotting the main paper data against communication rounds instead of total updates reveals the communication efficiency perspective: LE=16 converges in 125 rounds while LE=1 requires 2000 rounds, achieving

16\times

communication reduction with comparable final

g^{\text{STCH-Set}}

values.

	CIFAR-10			CIFAR-100			Medical
Method	Dir-10	Dir-20	Pat-10	Dir-10	Dir-20	Pat-20	Kvasir	FedISIC
FedAvg	0.981	0.981	0.982	0.995	0.982	0.977	0.999	0.945
FedProx	0.984	0.977	0.976	0.995	0.983	0.982	0.981	0.924
APFL	0.992	0.985	0.996	0.995	0.988	0.996	0.994	0.950
Ditto	0.984	0.982	0.997	0.994	0.987	0.996	0.994	0.951
FedRep	0.990	0.985	0.997	0.995	0.991	0.996	0.995	0.938

Appendix

Table 5: Jain’s Fairness Index (higher is better,

J=1

is perfect fairness).

	AG News			Kvasir
$K$	Min	Mean	Max	Min	Mean	Max
1	84.5	96.4	100.0	79.5	91.6	99.5
2	73.8	95.7	100.0	82.4	92.2	99.8
3	83.9	96.0	100.0	83.9	92.9	100.0
4	78.0	95.7	100.0	82.9	92.6	100.0
5	86.3	96.2	100.0	81.2	92.5	100.0

Appendix

Table 6: Test accuracy (%) vs. number of server models

K

on AG News and Kvasir datasets.

Federated Learning

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

See all 939 papers

Apr 13, 2026

Representation-Aligned Multi-Scale Personalization for Federated Learning

In federated learning (FL), accommodating clients with diverse resource constraints remains a significant challenge. A widely adopted approach is to use a shared full-size model, from which each client extracts a submodel aligned with its computational budget. However, regardless of the specific scoring strategy, these methods rely on the same global backbone, limiting both structural diversity and representational adaptation across clients. This paper presents FRAMP, a unified framework for personalized and resource-adaptive federated learning. Instead of relying on a fixed global model, FRAMP generates client-specific models from compact client descriptors, enabling fine-grained adaptation to both data characteristics and computational budgets. Each client trains a tailored lightweight submodel and aligns its learned representation with others to maintain global semantic consistency. Extensive experiments on vision and graph benchmarks demonstrate that FRAMP enhances generalization and adaptivity across a wide range of client settings.

Wenfei Liang, Wee Peng Tay

Department of Electrical and Electronic Engineering, Nanyang Technological University, Singapore.

Federated Learning Heterogeneous Federated Learning Representation Alignment

Ask PDF

May 26, 2026

Separate Aggregation of Split Network for Personalized Federated Learning

Federated learning enables collaborative model training without sharing raw data, but its performance can degrade substantially under heterogeneous client data distributions. A single global model often cannot satisfy diverse client requirements, so personalized federated learning has therefore been explored to improve client specific performance while preserving global generalization. Existing PFL methods often face a fundamental tradeoff in which stronger global sharing can undermine local specialization, whereas stronger local adaptation can lead to overfitting under limited data, label imbalance, and missing class scenarios. In this work, we propose PGFedSplit, a personalized federated learning framework that improves both personalization and global generalization under severe client heterogeneity. PGFedSplit adopts a split architecture and performs adaptive aggregation scheduling tailored to the roles of different model components, enabling stable knowledge sharing while maintaining client specific adaptation. Each client further leverages a mixture of locally extracted representations and synthetic representations generated from server side Gaussian statistics, improving robustness under label imbalance and missing class conditions. Extensive experiments on Fashion MNIST, CIFAR 10, CIFAR 100, and Tiny ImageNet demonstrate consistent improvements over state of the art PFL methods, with stable convergence and superior personalization in highly heterogeneous settings.

Yunseok Kang, Jaeyoung Song

Department of Electronics Engineering, Pusan National University, Republic of Korea

Federated Learning Aggregation Split Federated Learning Federated Learning Heterogeneous Federated Learning

Ask PDF

Mar 12, 2026

Personalized Federated Learning via Gaussian Generative Modeling

Federated learning has emerged as a paradigm to train models collaboratively on inherently distributed client data while safeguarding privacy. In this context, personalized federated learning tackles the challenge of data heterogeneity by equipping each client with a dedicated model. A prevalent strategy decouples the model into a shared feature extractor and a personalized classifier head, where the latter actively guides the representation learning. However, previous works have focused on classifier head-guided personalization, neglecting the potential personalized characteristics in the representation distribution. Building on this insight, we propose pFedGM, a method based on Gaussian generative modeling. The approach begins by training a Gaussian generator that models client heterogeneity via weighted re-sampling. A balance between global collaboration and personalization is then struck by employing a dual objective: a shared objective that maximizes inter-class distance across clients, and a local objective that minimizes intra-class distance within them. To achieve this, we decouple the conventional Gaussian classifier into a navigator for global optimization, and a statistic extractor for capturing distributional statistics. Inspired by the Kalman gain, the algorithm then employs a dual-scale fusion framework at global and local levels to equip each client with a personalized classifier head. In this framework, we model the global representation distribution as a prior and the client-specific data as the likelihood, enabling Bayesian inference for class probability estimation. The evaluation covers a comprehensive range of scenarios: heterogeneity in class counts, environmental corruption, and multiple benchmark datasets and configurations. pFedGM achieves superior or competitive performance compared to state-of-the-art methods.

Peng Hu, Jianwei Ma

P. Hu is with the School of Mathematics, Harbin Institute of Technology, Harbin, China. · J. Ma is with the Institute for Artificial Intelligence and the School of Earth and Space Sciences, Peking University, Beijing, China, and also with the Institute for Artificial Intelligence and the School of Mathematics, Harbin Institute of Technology, Harbin, China.

Federated Learning Heterogeneous Federated Learning Generative Modeling

Ask PDF

Abstract

Federated Learning

Explore similar work

Representation-Aligned Multi-Scale Personalization for Federated Learning

Separate Aggregation of Split Network for Personalized Federated Learning

Personalized Federated Learning via Gaussian Generative Modeling