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: ∙ pFedFDA [ 21 ] , ∙ FedLog [ 27 ] , ∙ pFedGM [ 28 ] , and ∙ 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 T communication rounds; MMS establishes a shared statistical basis in a single additional round, enabling SASP to construct personalized classifiers locally.
Method
Venue
CIFAR-10, Dir(0.3), 20 clients
CIFAR-10, Dir(0.1), 20 clients
local
global
unseen
rare
F@25
local
global
unseen
rare
F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]
AISTATS’17
60.77 ± 0.39
59.05 ± 0.98
58.16 ± 0.59
54.45 ± 1.29
40.40 ± 1.94
44.84 ± 0.52
45.01 ± 0.75
42.72 ± 0.95
50.32 ± 0.22
21.84 ± 0.92
FedRep [ 15 ]
ICML’21
79.25 ± 0.25
30.77 ± 0.09
0.00 ± 0.00
7.44 ± 0.20
0.00 ± 0.00
88.29 ± 0.07
18.35 ± 0.06
0.00 ± 0.00
13.20 ± 0.53
0.00 ± 0.00
FedROD [ 16 ]
ICLR’22
81.51 ± 0.09
59.94 ± 0.99
69.62 ± 1.03
52.60 ± 0.79
41.81 ± 0.58
89.26 ± 0.15
48.35 ± 0.62
50.47 ± 0.59
43.39 ± 0.81
22.73 ± 1.21
FedLC [ 17 ]
ICML’22
61.16 ± 0.28
59.76 ± 0.53
70.19 ± 0.58
52.05 ± 0.25
39.19 ± 0.32
44.88 ± 0.11
45.18 ± 0.46
47.80 ± 0.39
38.44 ± 0.87
10.48 ± 1.09
TABLE I : Comprehensive comparison on CIFAR-10 [ 1 ] under Dirichlet label heterogeneity. All methods are trained using ResNet-18. Symbols: ( † ) denotes the best, ( ∗ ) the second-best, and ( Δ ) the gain over the best baseline.
Method
Venue
CIFAR-100, Dir(0.3), 20 clients
CIFAR-100, Dir(0.1), 20 clients
local
global
unseen
rare
F@25
local
global
unseen
rare
F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]
AISTATS’17
35.37 ± 0.45
34.80 ± 0.82
33.33 ± 0.97
36.57 ± 1.28
23.00 ± 4.13
32.85 ± 0.27
32.73 ± 0.23
32.39 ± 0.23
28.85 ± 1.25
17.11 ± 1.83
FedRep [ 15 ]
ICML’21
37.74 ± 0.29
11.88 ± 0.17
0.00 ± 0.00
0.37 ± 0.15
0.00 ± 0.00
49.81 ± 0.31
9.33 ± 0.08
0.00 ± 0.00
0.23 ± 0.11
0.00 ± 0.00
FedROD [ 16 ]
ICLR’22
48.66 ± 0.12
34.72 ± 0.40
34.99 ± 0.51
35.10 ± 0.86
23.67 ± 1.77
56.97 ± 0.75 †
32.93 ± 0.69
33.39 ± 0.67
28.82 ± 0.45
18.22 ± 1.80
FedLC [ 17 ]
ICML’22
32.78 ± 0.17
32.04 ± 0.30
33.03 ± 0.10
31.53 ± 1.15
20.50 ± 1.53
30.78 ± 0.25
30.91 ± 0.16
31.73 ± 0.18
26.86 ± 0.36
16.72 ± 0.81
TABLE II : Comprehensive comparison on CIFAR-100 [ 1 ] under Dirichlet label heterogeneity. All methods are trained using ResNet-18. Symbols: ( † ) denotes the best, ( ∗ ) the second-best, and ( Δ ) the gain over the best baseline.
Method
Venue
Overall performance
Worst 25% of clients
local
global
unseen
rare
L@25
U@25
R@25
F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]
AISTATS’17
59.97 ± 0.28
61.27 ± 0.49
59.15 ± 0.93
64.32 ± 2.03
41.49 ± 3.59
40.56 ± 4.74
47.87 ± 2.98
37.10 ± 3.28
FedRep [ 15 ]
ICML’21
71.30 ± 0.46
25.19 ± 0.45
0.00 ± 0.00
18.93 ± 0.44
62.48 ± 0.42
0.00 ± 0.00
4.52 ± 0.07
0.00 ± 0.00
FedROD [ 16 ]
ICLR’22
71.62 ± 0.11
62.26 ± 2.19
63.13 ± 3.71
62.61 ± 0.72
61.63 ± 1.03
50.80 ± 4.48
49.37 ± 2.53
46.63 ± 3.26
FedLC [ 17 ]
ICML’22
57.10 ± 0.78
61.77 ± 0.22
63.21 ± 0.25
61.56 ± 1.55
39.74 ± 0.90
48.59 ± 0.86
47.95 ± 2.04
37.88 ± 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: ( † ) denotes the best, ( ∗ ) the second-best, and ( Δ ) the gain over the best baseline.
Method
State
local
global
unseen
rare
F@25
(a) CIFAR-100, Dir(0.3)
Before
39.08
38.77
36.94
40.57
24.02
46.20
25.58
14.04
18.72
1.33
FedETF [ 19 ]
After
( +7.12 )
( -13.19 )
( -22.90 )
( -21.85 )
( -22.69 )
Before
32.13
32.39
30.97
34.03
21.75
45.53
25.40
6.14
22.57
4.21
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 shrinkage
Local prior
Local mean
local
global
unseen
rare
L@25
U@25
R@25
F@25
(a) CIFAR-100, Dir(0.3), 20 clients
–
–
–
55.53
55.98
55.67
57.83
52.75
52.71
47.67
46.93
–
–
✓
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 backbone
Classifier head
CIFAR-10, Dir(0.3)
CIFAR-10, Dir(0.1)
CIFAR-100, Dir(0.3)
CIFAR-100, Dir(0.1)
FedRS
global
F@25
global
F@25
global
F@25
global
F@25
global
F@25
B0
H0
78 . 59
69 . 72
69 . 03
61 . 03
51 . 56
41 . 10
49 . 72
38 . 49
83 . 75
77 . 04
B0
H1
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 )
B1
H0
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
H1
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 I
-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. B0 : FedYoYo [ 26 ] ; B1 : the CGRL-trained FedGRC backbone. H0 : CCVR [ 29 ] on CIFAR-100 Dir(0.3), and FedCGS [ 32 ] elsewhere; H1 : MMS+SASP. The signed values in parentheses indicate gains over B0H0 ; bold entries identify full FedGRC. For either metric, Mij=M(Bi,Hj) , and the shaded row reports the excess gain I=M11−M10−M01+M00 .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Description
A. Representation
ResNet-18
Trained from scratch; no pretrained weights.
Λ={2,3,4}
Pooled dimensions: 128, 256, 512; concatenated dimension d=896 .
B. Consensus-Guided Representation Learning (CGRL)
λdir=λrad=1
Directional and magnitude guidance weights.
β=1
Shared-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.
Method
Venue
CIFAR-10, Dir(0.3), 20 clients
CIFAR-10, Dir(0.1), 20 clients
L@25
U@25
R@25
F@25
L@25
U@25
R@25
F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]
AISTATS’17
52.25 ± 0.25
47.32 ± 1.08
40.81 ± 2.39
40.40 ± 1.94
21.87 ± 0.96
35.88 ± 0.61
34.51 ± 0.08
21.84 ± 0.92
FedRep [ 15 ]
ICML’21
69.54 ± 0.51
0.00 ± 0.00
0.24 ± 0.10
0.00 ± 0.00
78.65 ± 0.56
0.00 ± 0.00
0.05 ± 0.04
0.00 ± 0.00
FedROD [ 16 ]
ICLR’22
72.43 ± 0.27
53.76 ± 0.64
43.20 ± 1.02
41.81 ± 0.58
79.87 ± 0.32
38.19 ± 0.63
23.87 ± 1.14
22.73 ± 1.21
FedLC [ 17 ]
ICML’22
47.91 ± 0.62
54.33 ± 0.80
41.37 ± 0.33
39.19 ± 0.32
15.95 ± 0.77
33.69 ± 0.54
15.79 ± 1.04
10.48 ± 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.
Method
Venue
CIFAR-100, Dir(0.3), 20 clients
CIFAR-100, Dir(0.1), 20 clients
L@25
U@25
R@25
F@25
L@25
U@25
R@25
F@25
Complete FL methods (method-specific training and classifiers)
FedAvg [ 14 ]
AISTATS’17
32.40 ± 0.99
30.05 ± 1.23
23.28 ± 4.25
23.00 ± 4.13
27.08 ± 0.15
30.65 ± 0.11
17.11 ± 1.83
17.11 ± 1.83
FedRep [ 15 ]
ICML’21
33.75 ± 0.68
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
44.22 ± 0.56
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
FedROD [ 16 ]
ICLR’22
44.88 ± 0.67
31.95 ± 0.76
23.67 ± 1.77
23.67 ± 1.77
52.59 ± 1.26 †
32.01 ± 0.95
18.22 ± 1.80
18.22 ± 1.80
FedLC [ 17 ]
ICML’22
29.55 ± 0.42
29.94 ± 0.04
20.50 ± 1.53
20.50 ± 1.53
23.55 ± 0.66
30.29 ± 0.09
16.89 ± 1.11
16.72 ± 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,Y
Client count, class count, and common federation-known label space.
D(k),Dc(k),Ic(k)
Client k ’s training set, class- c subset, and indices of that subset.
N(k),N;nc(k),Nc
Training-sample counts for each client and the federation, both overall and for class c .
θ=(ϕ,ψ)
Model parameters: backbone ϕ and classifier head ψ .
g,hϕ,qψ
Full model, feature extractor, and classifier head; g=qψ∘hϕ .
Appendix
TABLE XIII : Notation for the FedGRC framework, theoretical analysis, and experimental evaluation.
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
Statistical heterogeneity limits federated learning when a single global classifier cannot represent client-specific label distributions. In this work, we propose Personalized Federated Knowledge Distillation with Head Adaptation (pFedKDH), which aggregates only the shared backbone, keeps persistent client-specific heads, and uses a recalibrated global head as a teacher during local training. Across MNIST, Fashion-MNIST, CIFAR10, and CIFAR100 under class-wise Dirichlet partitions, pFedKDH obtains the best accuracy in most settings, with accuracy gaps up to 37.67% over the weakest baseline and consistently low standard deviation across repetitions. Component-wise diagnostics and convergence results support the role of persistent heads and distillation-guided local optimization under label-skewed data.
Polycarpo Souza Neto, José Mairton Barros da Silva Júnior, Charles Casimiro Cavalcante
Universidade Federal do Cear´a, Fortaleza, CE, Brazil · Uppsala University, Uppsala, Sweden
Personalized Federated Learning (PFL) aims to deliver effective client-specific models under heterogeneous distributions, yet existing methods suffer from shallow prototype alignment and brittle server-side distillation. We propose HEART-PFL, a dual-sided framework that (i) performs depth-aware Hierarchical Directional Alignment (HDA) using cosine similarity in the early stage and MSE matching in the deep stage to preserve client specificity, and (ii) stabilizes global updates through Adversarial Knowledge Transfer (AKT) with symmetric KL distillation on clean and adversarial proxy data. Using lightweight adapters with only 1.46M trainable parameters, HEART-PFL achieves state-of-the-art personalized accuracy on CIFAR-100, Flowers-102, and Caltech-101 (63.42%, 84.23%, and 95.67%, respectively) under Dirichlet non-IID partitions, and remains robust to out-of-domain proxy data. Ablation studies further confirm that HDA and AKT provide complementary gains in alignment, robustness, and optimization stability, offering insights into how the two components mutually reinforce effective personalization. Overall, these results demonstrate that HEART-PFL simultaneously enhances personalization and global stability, highlighting its potential as a strong and scalable solution for PFL(code available at https://github.com/danny0628/HEART-PFL).