cs.LGOct 22, 2026

Is High Local Accuracy Enough? Geometric Consensus with Support-Adaptive Statistical Personalization for Federated Learning

Authors: Junyu Lin, Bosong Qi, Pengcheng Shi, Jiayuan Li

Organizations: School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430072, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Real-world deployment of personalized federated learning (PFL) requires adapting to heterogeneous clients while preserving recognition performance beyond local distributions. However, high local accuracy can conceal failures on locally rare or absent classes. Effective personalization should therefore strengthen global recognition while raising the personalization floor. We propose an evaluation protocol that complements local accuracy with global, rare-class, and unseen-class evaluations to reveal generalization obscured by local scores, and assesses personalization floors among the worst-performing clients to expose weaknesses hidden by aggregate performance. To meet these deployment challenges, we propose FedGRC, a unified PFL framework integrating generalization and personalization through (i) Consensus-Guided Representation Learning (CGRL), which learns discriminative representations through shared directional and magnitude guidance; (ii) Multilayer Metric Synthesis (MMS), which constructs a shared covariance metric from these multilayer representations for statistical classification; and (iii) Support-Adaptive Shrinkage Personalization (SASP), which adapts class means within this metric by shrinking local estimates toward shared means according to local sample support, enabling adaptation while retaining shared knowledge. Extensive experiments, including benchmarking against 21 methods across 5 heterogeneity settings on 3 datasets and comprehensive ablation studies, validate the effectiveness of FedGRC. On CIFAR-10 and CIFAR-100, FedGRC surpasses state-of-the-art (SOTA) baselines in global recognition and personalization floors by at least 2.17% and 2.13%, respectively. On the real-world heterogeneous FedRS dataset, it outperforms the respective best baselines by 2.15%, 3.30%, and 4.61% in local adaptation, global recognition, and the personalization floor.

Figures & tables

Fig. 1 : Local personalization and global generalization on CIFAR-10 and CIFAR-100. Colors identify methods: ∙\bullet pFedFDA [ 21 ] , ∙\bullet FedLog [ 27 ] , ∙\bullet pFedGM [ 28 ] , and ∙\bullet our FedGRC. Dir(0.3) and Dir(0.1) denote Dirichlet label partitions with strong and extreme heterogeneity, respectively. Local denotes accuracy on local test data; global denotes accuracy on a class-balanced test set covering all classes.
Fig. 2 : Overview of FedGRC. CGRL learns shared representations over TT communication rounds; MMS establishes a shared statistical basis in a single additional round, enabling SASP to construct personalized classifiers locally.
MethodVenueCIFAR-10, Dir(0.3), 20 clients CIFAR-10, Dir(0.1), 20 clients
localglobalunseenrareF@25 localglobalunseenrareF@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]AISTATS’1760.77 ±\pm 0.3959.05 ±\pm 0.9858.16 ±\pm 0.5954.45 ±\pm 1.2940.40 ±\pm 1.94 44.84 ±\pm 0.5245.01 ±\pm 0.7542.72 ±\pm 0.9550.32 ±\pm 0.2221.84 ±\pm 0.92
FedRep [ 15 ]ICML’2179.25 ±\pm 0.2530.77 ±\pm 0.090.00 ±\pm 0.007.44 ±\pm 0.200.00 ±\pm 0.00 88.29 ±\pm 0.0718.35 ±\pm 0.060.00 ±\pm 0.0013.20 ±\pm 0.530.00 ±\pm 0.00
FedROD [ 16 ]ICLR’2281.51 ±\pm 0.0959.94 ±\pm 0.9969.62 ±\pm 1.0352.60 ±\pm 0.7941.81 ±\pm 0.58 89.26 ±\pm 0.1548.35 ±\pm 0.6250.47 ±\pm 0.5943.39 ±\pm 0.8122.73 ±\pm 1.21
FedLC [ 17 ]ICML’2261.16 ±\pm 0.2859.76 ±\pm 0.5370.19 ±\pm 0.5852.05 ±\pm 0.2539.19 ±\pm 0.32 44.88 ±\pm 0.1145.18 ±\pm 0.4647.80 ±\pm 0.3938.44 ±\pm 0.8710.48 ±\pm 1.09
TABLE I : Comprehensive comparison on CIFAR-10 [ 1 ] under Dirichlet label heterogeneity. All methods are trained using ResNet-18. Symbols: ( †\dagger ) denotes the best, ( ∗* ) the second-best, and ( Δ\Delta ) the gain over the best baseline.
MethodVenueCIFAR-100, Dir(0.3), 20 clients CIFAR-100, Dir(0.1), 20 clients
localglobalunseenrareF@25 localglobalunseenrareF@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]AISTATS’1735.37 ±\pm 0.4534.80 ±\pm 0.8233.33 ±\pm 0.9736.57 ±\pm 1.2823.00 ±\pm 4.13 32.85 ±\pm 0.2732.73 ±\pm 0.2332.39 ±\pm 0.2328.85 ±\pm 1.2517.11 ±\pm 1.83
FedRep [ 15 ]ICML’2137.74 ±\pm 0.2911.88 ±\pm 0.170.00 ±\pm 0.000.37 ±\pm 0.150.00 ±\pm 0.00 49.81 ±\pm 0.319.33 ±\pm 0.080.00 ±\pm 0.000.23 ±\pm 0.110.00 ±\pm 0.00
FedROD [ 16 ]ICLR’2248.66 ±\pm 0.1234.72 ±\pm 0.4034.99 ±\pm 0.5135.10 ±\pm 0.8623.67 ±\pm 1.77 56.97 ±\pm 0.75 †32.93 ±\pm 0.6933.39 ±\pm 0.6728.82 ±\pm 0.4518.22 ±\pm 1.80
FedLC [ 17 ]ICML’2232.78 ±\pm 0.1732.04 ±\pm 0.3033.03 ±\pm 0.1031.53 ±\pm 1.1520.50 ±\pm 1.53 30.78 ±\pm 0.2530.91 ±\pm 0.1631.73 ±\pm 0.1826.86 ±\pm 0.3616.72 ±\pm 0.81
TABLE II : Comprehensive comparison on CIFAR-100 [ 1 ] under Dirichlet label heterogeneity. All methods are trained using ResNet-18. Symbols: ( †\dagger ) denotes the best, ( ∗* ) the second-best, and ( Δ\Delta ) the gain over the best baseline.
MethodVenueOverall performance Worst 25% of clients
localglobalunseenrare L@25U@25R@25F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]AISTATS’1759.97 ±\pm 0.2861.27 ±\pm 0.4959.15 ±\pm 0.9364.32 ±\pm 2.03 41.49 ±\pm 3.5940.56 ±\pm 4.7447.87 ±\pm 2.9837.10 ±\pm 3.28
FedRep [ 15 ]ICML’2171.30 ±\pm 0.4625.19 ±\pm 0.450.00 ±\pm 0.0018.93 ±\pm 0.44 62.48 ±\pm 0.420.00 ±\pm 0.004.52 ±\pm 0.070.00 ±\pm 0.00
FedROD [ 16 ]ICLR’2271.62 ±\pm 0.1162.26 ±\pm 2.1963.13 ±\pm 3.7162.61 ±\pm 0.72 61.63 ±\pm 1.0350.80 ±\pm 4.4849.37 ±\pm 2.5346.63 ±\pm 3.26
FedLC [ 17 ]ICML’2257.10 ±\pm 0.7861.77 ±\pm 0.2263.21 ±\pm 0.2561.56 ±\pm 1.55 39.74 ±\pm 0.9048.59 ±\pm 0.8647.95 ±\pm 2.0437.88 ±\pm 0.49
TABLE III : Comprehensive comparison on FedRS [ 3 ] with 8 remote-sensing sources, 15 classes, and 135 clients. All methods are trained using ResNet-18. Symbols: ( †\dagger ) denotes the best, ( ∗* ) the second-best, and ( Δ\Delta ) the gain over the best baseline.
MethodStatelocalglobalunseenrareF@25
(a) CIFAR-100, Dir(0.3)
 Before39.0838.7736.9440.5724.02
  46.2025.5814.0418.721.33
FedETF [ 19 ]After( +7.12 )( -13.19 )( -22.90 )( -21.85 )( -22.69 )
 Before32.1332.3930.9734.0321.75
  45.5325.406.1422.574.21
TABLE IV : Recognition before and after local personalization. Before: shared predictors. After: FedETF [ 19 ] , full-model adaptation; pFedGM [ 28 ] , classifier adaptation with the generator unchanged; FedGRC, SASP with CGRL+MMS unchanged. The signed values in parentheses indicate changes (After minus Before).
Table 7
Count shrinkageLocal priorLocal meanlocalglobalunseenrareL@25U@25R@25F@25
(a) CIFAR-100, Dir(0.3), 20 clients
–––55.5355.9855.6757.8352.7552.7147.6746.93
––✓57.30 ( ++ 1.77 )45.17 ( −- 10.81 )61.88 ( ++ 6.21 )1.17 ( −- 56.66 )54.47 ( ++ 1.72 )59.38 ( ++ 6.67 )0.00 ( −- 47.67 )0.00 ( −- 46.93 )
–✓✓58.91 ( ++ 3.38 )45.29 ( −- 10.69 )60.66 ( ++ 4.99 )1.21 ( −- 56.62 )56.19 ( ++ 3.44 )58.24 ( ++ 5.53 )0.00 ( −- 47.67 )0.00 ( −- 46.93 )
✓✓✓57.21 ( ++ 1.68 )55.63 ( −- 0.35 )54.42 ( −- 1.25 )57.04 ( −- 0.79 )54.30 ( ++ 1.55 )51.31 ( −- 1.40 )45.67 ( −- 2.00 )44.33 ( −- 2.60 )
✓–✓55.61 ( ++ 0.08 )55.71 ( −- 0.27 )55.75 ( ++ 0.08 )58.00 ( ++ 0.17 )52.68 ( −- 0.07 )52.90 ( ++ 0.19 )48.33 ( ++ 0.66 )47.56 ( ++ 0.63 )
TABLE VIII : Ablation of local means, count-based shrinkage, and local class priors in SASP. (✓) denotes inclusion. All configurations use the same CGRL backbone and MMS readout. Locally absent classes retain shared means; without local priors, class priors are equal. The signed values in parentheses indicate changes relative to the shared-mean classifier.
Training backboneClassifier headCIFAR-10, Dir(0.3)CIFAR-10, Dir(0.1)CIFAR-100, Dir(0.3)CIFAR-100, Dir(0.1)FedRS
globalF@25globalF@25globalF@25globalF@25globalF@25
B0B_{0}H0H_{0}78 . 5969 . 7269 . 0361 . 0351 . 5641 . 1049 . 7238 . 4983 . 7577 . 04
B0B_{0}H1H_{1}78 . 55 ( −- 0.04 )69 . 32 ( −- 0.40 )69 . 84 ( ++ 0.81 )60 . 38 ( −- 0.65 )51 . 05 ( −- 0.51 )40 . 98 ( −- 0.12 )50 . 00 ( ++ 0.28 )37 . 61 ( −- 0.88 )84 . 65 ( ++ 0.90 )79 . 57 ( ++ 2.53 )
B1B_{1}H0H_{0}81 . 17 ( ++ 2.58 )71 . 39 ( ++ 1.67 )72 . 09 ( ++ 3.06 )63 . 47 ( ++ 2.44 )54 . 45 ( ++ 2.89 )45 . 07 ( ++ 3.97 )52 . 08 ( ++ 2.36 )41 . 07 ( ++ 2.58 )85 . 69 ( ++ 1.94 )76 . 88 ( −- 0.16 )
B1\bm{B_{1}}H1\bm{H_{1}}80 . 76 ( ++ 2.17 )71 . 85 ( ++ 2.13 )73 . 05 ( ++ 4.02 )63 . 85 ( ++ 2.82 )55 . 71 ( ++ 4.15 )47 . 56 ( ++ 6.46 )53 . 22 ( ++ 3.50 )41 . 83 ( ++ 3.34 )87 . 08 ( ++ 3.33 )81 . 65 ( ++ 4.61 )
Excess gain II-0 . 37+0 . 86+0 . 15+1 . 03+1 . 77+2 . 61+0 . 86+1 . 64+0 . 49+2 . 24
TABLE IX : Backbone–classifier cross-comparison against strong baseline combinations across all heterogeneity settings. B0B_{0} : FedYoYo [ 26 ] ; B1B_{1} : the CGRL-trained FedGRC backbone. H0H_{0} : CCVR [ 29 ] on CIFAR-100 Dir(0.3), and FedCGS [ 32 ] elsewhere; H1H_{1} : MMS+SASP. The signed values in parentheses indicate gains over B0H0B_{0}H_{0} ; bold entries identify full FedGRC. For either metric, Mij=M(Bi,Hj)M_{ij}=M(B_{i},H_{j}) , and the shaded row reports the excess gain I=M11−M10−M01+M00I=M_{11}-M_{10}-M_{01}+M_{00} .
Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

ConfigurationDescription
A. Representation
ResNet-18Trained from scratch; no pretrained weights.
Λ={2,3,4}\Lambda=\{2,3,4\}Pooled dimensions: 128, 256, 512; concatenated dimension d=896d=896 .
B. Consensus-Guided Representation Learning (CGRL)
λdir=λrad=1\lambda_{\mathrm{dir}}=\lambda_{\mathrm{rad}}=1Directional and magnitude guidance weights.
β=1\beta=1Shared-prototype weight in the directional loss.
Appendix
TABLE X : Key parameters and implementation details of FedGRC.
Fig. 3 : Heterogeneity in client–class distributions of training and local test samples on CIFAR-10 [ 1 ] under Dir(0.3) and Dir(0.1). Blank entries indicate zero samples.
Fig. 4 : Heterogeneity in client–class distributions of training and local test samples on CIFAR-100 [ 1 ] under Dir(0.3) and Dir(0.1). Blank entries indicate zero samples.
Fig. 5 : Heterogeneity in client–class distributions of training and local test samples on FedRS [ 3 ] under NIID-1. Blank entries indicate zero samples.
MethodVenueCIFAR-10, Dir(0.3), 20 clients CIFAR-10, Dir(0.1), 20 clients
L@25U@25R@25F@25 L@25U@25R@25F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]AISTATS’1752.25 ±\pm 0.2547.32 ±\pm 1.0840.81 ±\pm 2.3940.40 ±\pm 1.94 21.87 ±\pm 0.9635.88 ±\pm 0.6134.51 ±\pm 0.0821.84 ±\pm 0.92
FedRep [ 15 ]ICML’2169.54 ±\pm 0.510.00 ±\pm 0.000.24 ±\pm 0.100.00 ±\pm 0.00 78.65 ±\pm 0.560.00 ±\pm 0.000.05 ±\pm 0.040.00 ±\pm 0.00
FedROD [ 16 ]ICLR’2272.43 ±\pm 0.2753.76 ±\pm 0.6443.20 ±\pm 1.0241.81 ±\pm 0.58 79.87 ±\pm 0.3238.19 ±\pm 0.6323.87 ±\pm 1.1422.73 ±\pm 1.21
FedLC [ 17 ]ICML’2247.91 ±\pm 0.6254.33 ±\pm 0.8041.37 ±\pm 0.3339.19 ±\pm 0.32 15.95 ±\pm 0.7733.69 ±\pm 0.5415.79 ±\pm 1.0410.48 ±\pm 1.09
Appendix
TABLE XI : Client recognition floors on CIFAR-10 [ 1 ] . L@25, U@25, R@25, and F@25 average the lowest 25% (5 of 20 clients), with clients ranked separately for each metric.
MethodVenueCIFAR-100, Dir(0.3), 20 clients CIFAR-100, Dir(0.1), 20 clients
L@25U@25R@25F@25 L@25U@25R@25F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]AISTATS’1732.40 ±\pm 0.9930.05 ±\pm 1.2323.28 ±\pm 4.2523.00 ±\pm 4.13 27.08 ±\pm 0.1530.65 ±\pm 0.1117.11 ±\pm 1.8317.11 ±\pm 1.83
FedRep [ 15 ]ICML’2133.75 ±\pm 0.680.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00 44.22 ±\pm 0.560.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
FedROD [ 16 ]ICLR’2244.88 ±\pm 0.6731.95 ±\pm 0.7623.67 ±\pm 1.7723.67 ±\pm 1.77 52.59 ±\pm 1.26 †32.01 ±\pm 0.9518.22 ±\pm 1.8018.22 ±\pm 1.80
FedLC [ 17 ]ICML’2229.55 ±\pm 0.4229.94 ±\pm 0.0420.50 ±\pm 1.5320.50 ±\pm 1.53 23.55 ±\pm 0.6630.29 ±\pm 0.0916.89 ±\pm 1.1116.72 ±\pm 0.81
Appendix
TABLE XII : Client recognition floors on CIFAR-100 [ 1 ] . L@25, U@25, R@25, and F@25 average the lowest 25% (5 of 20 clients), with clients ranked separately for each metric.
Symbol(s)Definition
A. Federated Learning Problem Formulation
K,C,YK,C,\mathcal{Y}Client count, class count, and common federation-known label space.
D(k),Dc(k),Ic(k)\mathcal{D}^{(k)},\mathcal{D}_{c}^{(k)},\mathcal{I}_{c}^{(k)}Client kk ’s training set, class- cc subset, and indices of that subset.
N(k),N; nc(k),NcN^{(k)},N;\ n_{c}^{(k)},N_{c}Training-sample counts for each client and the federation, both overall and for class cc .
θ=(ϕ,ψ)\bm{\theta}=(\bm{\phi},\bm{\psi})Model parameters: backbone ϕ\bm{\phi} and classifier head ψ\bm{\psi} .
g,hϕ,qψg,h_{\bm{\phi}},q_{\bm{\psi}}Full model, feature extractor, and classifier head; g=qψ∘hϕg=q_{\bm{\psi}}\circ h_{\bm{\phi}} .
Appendix
TABLE XIII : Notation for the FedGRC framework, theoretical analysis, and experimental evaluation.
More in this topic

Federated Learning

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

See all 951 papers

Explore similar work

CardsList

  1. Separate Aggregation of Split Network for Personalized Federated Learning

    May 26, 2026Yunseok Kang, Jaeyoung SongFederated Learning AggregationSplit Federated Learning

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

  2. Personalized Federated Learning through Global Knowledge Distillation and Local Head Adaptation

    Sep 15, 2026Polycarpo Souza Neto, José Mairton Barros da Silva Júnior, Charles Casimiro CavalcanteFederated Knowledge DistillationNon-IID Federated Learning

    Universidade Federal do Cear´a, Fortaleza, CE, Brazil · Uppsala University, Uppsala, Sweden

  3. HEART-PFL: Stable Personalized Federated Learning under Heterogeneity with Hierarchical Directional Alignment and Adversarial Knowledge Transfer

    Mar 25, 2026Minjun Kim, Minje KimFederated LearningHeterogeneous Federated Learning

    Promedius Inc.