Personalized federated learning must decide which clients should share representations. We introduce HDC-FL (Head-Derived Clustering for Personalized Federated Learning), which reads that decision off a signal shared-encoder training already produces: the client-specific classifi- er heads. At periodic refresh rounds the server projects the cached heads into a normalized low-rank descriptor space, groups the descriptors, builds one encoder prototype per group from the cached client encoders, and returns to each client a soft mixture of those prototypes paired with its own unchanged head. Client-specific heads are never averaged or distributed to oth- er clients; the shared-encoder trajectory is left intact; and no second optimization trajectory, additional local objective, or post-training fine-tuning stage is required. Across 11 CIFAR-10 and CIFAR-100 configurations covering label-subset, Dirichlet and rotation heterogeneity with ResNet-18 and MobileNetV2 encoders, HDC-FL is competitive with the strongest personalized and clustered baselines in macro-averaged mean accuracy and attains the numerically highest macro-averaged worst-client and bottom-20% accuracy among the evaluated methods. The tail results distinguish methods that look similar when only their means are compared. A controlled donor-swap diagnostic further finds positive donor-ranking associations between head similarity and encoder compatibility in all three settings tested, most clearly under covariate shift.
Figures & tables
Figure 1: HDC-FL overview. At each refresh the server converts cached client-specific heads into low-rank client descriptors, groups them, constructs group-level encoder prototypes from cached client encoders, and returns a soft prototype mixture to each client. The mixture is paired with the client’s unchanged private head, and the shared training trajectory continues unmodified.
Method
C10 LS2
C10 LS3
C10 Rot
C100 α=0.07
C100 α=0.3
C100 LS2
C100 LS3
C100 Rot
Avg. Rank ↓
FedAvg-SE
97.17 ± 0.62
95.51 ± 0.71
86.42 ± 0.80
81.15 ± 1.06
72.09 ± 0.34
81.23 ± 0.63
76.88 ± 0.17
57.42 ± 0.68
7.25
FedBABU
96.99 ± 0.63
97.18 ± 1.35
90.05 ± 0.43
87.73 ± 1.51
82.19 ± 0.16
84.81 ± 4.02
82.33 ± 3.92
68.54 ± 1.91
2.12
FedALA
96.51 ± 0.65
95.50 ± 0.85
86.73 ± 0.03
81.45 ± 1.39
74.67 ± 0.03
80.96 ± 0.14
77.75 ± 0.40
63.86 ± 0.21
6.88
FedCE
96.89 ± 0.02
95.92 ± 0.12
86.52 ± 0.28
82.22 ± 0.41
77.19 ± 0.06
82.29 ± 0.76
78.98 ± 0.45
64.37 ± 0.26
4.50
FedSoft
96.40 ± 0.09
95.44 ± 1.39
84.41 ± 0.48
81.90 ± 0.94
75.41 ± 0.22
81.63 ± 0.25
79.23 ± 0.04
61.06 ± 0.59
6.88
IFCA
96.36 ± 1.05
94.38 ± 0.42
87.18 ± 0.13
80.12 ± 0.28
75.93 ± 0.55
80.61 ± 1.00
78.40 ± 0.23
63.69 ± 0.46
7.62
Table 1: Mean client test accuracy (%) on ResNet-18, mean ± sample standard deviation over three seeds. Average rank is computed within each setting and then averaged over the eight settings. Bold and underline mark the best and second-best entry per setting.
Method
Mean
Worst
Bot.20%
Gap
Tail rank
FedAvg-SE
83.35
76.95
78.95
11.15
6.00
FedBABU
87.58
55.32
73.78
38.37
8.18
FedALA
84.19
77.68
79.75
11.07
5.73
FedCE
85.03
78.73
80.69
10.94
3.73
FedSoft
84.15
76.52
79.31
12.38
6.00
IFCA
83.64
71.25
77.58
17.90
7.27
Table 2: Client-level results macro-averaged over all 11 configurations.
Variant
Mean
Worst
Bot.20%
No prototype
77.57 ± 0.40
69.85 ± 3.97
72.20 ± 1.46
One prototype
80.30 ± 0.34
72.05 ± 2.90
74.93 ± 1.16
Head-derived, G=4
81.04 ± 0.73
72.57 ± 1.67
75.47 ± 1.10
Table 3: Controlled grouping ablation on CIFAR-10 rotation without pretrained initialization. Here G=4 is fixed to match the four known rotation conditions and isolate the effect of head-derived grouping. Values are mean ± sample standard deviation over three seeds.
C10 LS3
C100 LS3
C100 Rot
Spearman ρ (loss)
0.193
0.117
0.243
AUROC (loss)
0.665
0.565
0.783
Beneficial-donor rate
0.156
0.711
0.245
AP
0.513
0.782
0.642
AP lift over prevalence
0.357
0.071
0.397
Top-1 − random (pp)
1.57
0.41
2.50
Table 4: Donor-compatibility diagnostic on MobileNetV2, averaged over targets and seeds. Spearman ρ measures rank association between head similarity and loss benefit, while AUROC and average precision (AP) measure discrimination between beneficial and non-beneficial donors. The beneficial-donor rate gives the positive-class prevalence against which AP should be interpreted. The final two rows report the accuracy gain from selecting the nearest-head donor rather than a random donor, and from the closest head-distance quintile relative to the farthest.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Method-specific configuration
FedAvg-SE
The encoder is aggregated across participating clients using sample-count weighting, while each client retains its own classifier head. No additional personalization stage is applied.
FedBABU
The classifier is kept fixed during federated training and only the encoder/body is locally optimized and aggregated. Final personalization uses 5 epochs of full-model fine-tuning with SGD, learning rate 10−2 , momentum 0 , and weight decay 0 .
FedALA
ALA depth p=2 , local-data fraction s=80% , and weight-learning rate η=1 . Initial ALA weight optimization is capped at 50 epochs with patience 3 and tolerance 10−4 ; subsequent rounds use one ALA weight-update epoch.
IFCA
Uses S=4 cluster models. Client-to-cluster assignment is based on model loss evaluated on at most 512 local samples.
PACFL
Client subspace dimension p=3 , hierarchical-clustering threshold 1.5 radians, average linkage, and one-shot client clustering.
Table 5: Method-specific baseline configurations used in our experiments. FedAvg-SE denotes FedAvgSharedEncoder.
Method
CIFAR-100 LS3
CIFAR-10 α=0.07
CIFAR-10 LS3
Avg. Rank ↓
FedAvg-SE
77.43 ± 0.37
95.85 ± 0.28
95.25 ± 0.74
6.67
FedBABU
81.27 ± 1.89
96.26 ± 1.19
96.05 ± 1.42
2.33
FedALA
77.77 ± 0.27
95.99 ± 0.10
94.87 ± 1.17
5.67
FedCE
78.87 ± 0.43
96.12 ± 0.38
95.96 ± 0.51
4.00
FedSoft
79.57 ± 0.48
95.07 ± 1.13
95.52 ± 0.39
5.33
IFCA
77.69 ± 0.49
90.90 ± 1.54
94.83 ± 0.65
8.67
Appendix
Table 6: MobileNetV2 mean client test accuracy (%), mean ± sample standard deviation over seeds 42, 123 and 2026. Bold and underline mark the best and second-best entry per setting.
Method
C10 LS2
C10 LS3
C10 Rot
C100 α=.07
C100 α=.3
C100 LS2
C100 LS3
C100 Rot
M-C100 LS3
M-C10 α=.07
M-C10 LS3
FedAvg-SE
88.73 ± 5.75
90.23 ± 1.03
79.53 ± 4.86
74.29 ± 0.82
67.10 ± 0.69
75.70 ± 0.69
68.06 ± 1.50
52.29 ± 1.28
72.45 ± 0.60
87.92 ± 2.67
90.17 ± 0.18
FedBABU
77.64 ± 10.88
82.66 ± 13.87
75.68 ± 3.07
55.35 ± 26.75
34.07 ± 6.89
51.94 ± 31.17
35.67 ± 5.29
20.04 ± 2.68
27.40 ± 6.79
76.28 ± 24.78
71.84 ± 18.52
FedALA
86.89 ± 7.23
90.20 ± 2.12
80.41 ± 4.74
72.97 ± 1.82
68.51 ± 1.17
74.21 ± 2.59
72.82 ± 0.84
57.50 ± 0.26
72.57 ± 0.79
88.87 ± 0.48
89.53 ± 2.33
FedCE
88.33 ± 5.50
91.50 ± 0.28
79.46 ± 4.10
75.33 ± 1.96
72.30 ± 1.19
75.92 ± 0.07
72.29 ± 2.25
57.76 ± 0.76
73.53 ± 1.19
88.28 ± 2.07
91.36 ± 0.79
FedSoft
87.55 ± 4.39
87.73 ± 3.01
75.76 ± 4.87
76.05 ± 0.55
66.92 ± 0.58
75.64 ± 0.14
72.14 ± 0.69
54.98 ± 1.36
73.90 ± 1.97
82.63 ± 6.61
88.40 ± 2.80
IFCA
86.15 ± 8.06
69.67 ± 26.74
81.83 ± 2.86
73.27 ± 0.01
69.68 ± 0.27
73.11 ± 0.49
68.78 ± 8.28
56.89 ± 3.31
70.65 ± 2.51
47.28 ± 13.44
86.41 ± 4.12
Appendix
Table 7: Final worst-client test accuracy (%), mean ± sample standard deviation over three seeds.
Method
C10 LS2
C10 LS3
C10 Rot
C100 α=.07
C100 α=.3
C100 LS2
C100 LS3
C100 Rot
M-C100 LS3
M-C10 α=.07
M-C10 LS3
FedAvg-SE
92.05 ± 1.36
91.84 ± 0.71
82.45 ± 2.19
75.96 ± 1.24
68.10 ± 0.16
77.20 ± 0.39
71.29 ± 0.29
53.68 ± 1.77
73.74 ± 0.20
90.26 ± 1.23
91.91 ± 0.54
FedBABU
89.45 ± 4.26
91.59 ± 5.49
82.78 ± 1.30
75.71 ± 10.75
68.89 ± 0.82
64.71 ± 20.65
59.18 ± 18.77
47.08 ± 10.85
56.46 ± 9.34
88.73 ± 5.46
87.03 ± 7.76
FedALA
91.64 ± 2.19
91.58 ± 2.75
82.78 ± 1.97
74.92 ± 1.38
69.90 ± 1.20
76.92 ± 1.30
73.85 ± 0.01
59.89 ± 0.10
74.27 ± 1.04
90.86 ± 0.99
90.67 ± 1.95
FedCE
92.11 ± 1.98
92.55 ± 0.37
81.61 ± 3.01
77.17 ± 0.79
73.39 ± 1.52
77.74 ± 0.62
74.15 ± 1.02
59.88 ± 0.18
74.99 ± 1.12
91.00 ± 1.45
92.94 ± 0.57
FedSoft
90.48 ± 1.67
91.20 ± 2.26
78.41 ± 2.87
77.01 ± 0.84
70.82 ± 1.10
77.36 ± 0.18
74.41 ± 0.36
56.73 ± 1.22
75.73 ± 0.92
88.32 ± 3.37
91.91 ± 1.09
IFCA
90.35 ± 4.31
86.74 ± 5.44
84.12 ± 1.78
75.84 ± 0.04
71.63 ± 0.09
74.58 ± 1.31
73.84 ± 2.27
59.78 ± 0.73
72.80 ± 0.63
73.45 ± 1.85
90.31 ± 0.43
Appendix
Table 8: Final bottom-20% client test accuracy (%), mean ± sample standard deviation over three seeds.
Method
C10 LS2
C10 LS3
C10 Rot
C100 α=.07
C100 α=.3
C100 LS2
C100 LS3
C100 Rot
M-C100 LS3
M-C10 α=.07
M-C10 LS3
Macro
FedAvg-SE
11.17
8.70
11.45
13.79
10.06
11.90
15.26
9.89
9.62
12.08
8.69
11.15
FedBABU
22.36
17.24
18.70
38.38
54.44
42.58
55.82
58.12
62.91
23.72
27.78
38.37
FedALA
12.84
8.05
11.55
16.47
11.59
12.13
9.16
10.51
9.35
11.13
8.95
11.07
FedCE
11.67
7.87
13.39
13.81
8.57
12.67
11.21
11.86
10.05
11.72
7.48
10.94
FedSoft
12.34
11.29
14.03
12.94
12.23
12.23
12.06
10.65
10.72
17.37
10.31
12.38
IFCA
13.54
28.94
10.52
13.26
10.80
14.32
14.13
12.46
14.31
52.72
11.91
17.90
Appendix
Table 9: Final fairness gap (best minus worst client accuracy, percentage points; lower is better). Each setting is first averaged over its seeds; the last column macro-averages the 11 setting-level means.
Figure 2: Mean client test accuracy over communication rounds for CIFAR-10 rotation, CIFAR-100 LS3 and CIFAR-100 α=0.3 , averaged over three seeds. Before activation the HDC-FL curve reports the shared trajectory; afterwards it reports the personalized prototype mixtures constructed from it.
Figure 3: Learning curves for the configurations not shown in Figure 2 , averaged over three seeds. Shaded regions are one sample standard deviation.
Figure 4: Grouping diagnostics over communication rounds on CIFAR-10 rotation without pretrained initialization with four groups: cluster separation (silhouette), agreement between consecutive assignments (ARI), membership confidence (entropy), and personalization gain. Lines and shaded regions are the mean and one standard deviation over three seeds.
Descriptor
Mean
Sil.
ARI
Raw head
85.27 ± 0.09
0.027 ± 0.002
0.152 ± 0.065
Low rank ( r=8 )
85.24 ± 0.31
0.108 ± 0.012
0.401 ± 0.080
Appendix
Table 10: Descriptor. Sil. is the silhouette score of the resulting clustering and ARI the agreement between the assignments of consecutive refreshes; both are higher-is-better. Means ± sample standard deviations over three seeds.
Rank r
Mean
Worst
Bot.20%
Silhouette ↑
8
40.85
35.57
36.71
0.195
20
40.96
35.38
36.36
0.126
40
40.24
34.02
35.91
0.090
60
40.50
34.05
35.71
0.068
80
40.23
34.39
35.84
0.059
Appendix
Table 11: Descriptor-rank sensitivity, single seed.
Membership
Mean
Worst
Bot.20%
Hard assignment
78.89 ± 0.46
73.55 ± 1.18
75.39 ± 0.52
Soft mixture
79.74 ± 0.40
74.63 ± 0.26
76.34 ± 0.46
Appendix
Table 12: Prototype membership. Means ± sample standard deviations over three seeds.
Table 14: Accuracy gain (percentage points) of a single donor-encoder swap, by low-rank head-distance quintile; Q1 is the closest fifth of donors. Values are means over the 60 target × seed units of a setting, with standard errors.
C10 LS3
C100 LS3
C100 Rot
Spearman ρ , low rank
0.193
0.117
0.243
Spearman ρ , raw
0.199
0.050
0.288
Difference
−0.006
+0.067
−0.045
AUROC, low rank
0.665
0.565
0.783
AUROC, raw
0.677
0.502
0.793
Difference
−0.012
+0.064
−0.010
Appendix
Table 15: Low-rank versus raw head similarity as a donor-ranking signal. Positive differences favour the low-rank descriptor.
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
Hierarchical federated learning (HFL) leverages edge servers for partial aggregation in edge computing. Yet existing FL methods lack mechanisms for jointly optimizing cluster assignment and client selection under data heterogeneity. This paper proposes Fed-BAC, which integrates additive cluster personalization with a two-level bandit framework: contextual bandits at the cloud learn server-to-cluster assignments, while Thompson Sampling at each edge server identifies high-contributing clients. The additive decomposition enables the sharing of knowledge between groups through a globally aggregated network, while cluster-specific networks capture distribution variations. Across three classification benchmarks (CIFAR-10, SVHN, Fashion-MNIST) under moderate (α=0.5) and severe (α=0.1) Dirichlet non-IID partitioning, Fed-BAC achieves distributed accuracy gains of up to +35.5pp over HierFAVG and +8.4pp over IFCA, while requiring only 80% client participation, converging 1.5 to 4.8× faster depending on dataset and accuracy target, and improving cross-server fairness. These gains are further validated at 5× deployment scale on CIFAR-10. The advantage of Fed-BAC increases with heterogeneity severity, confirming that additive cluster personalization becomes increasingly valuable as data distributions diverge.
Satwat Bashir, Tasos Dagiuklas, Muddesar Iqbal
Computer Science and Informatics · London South Bank University · London, UK +1
Statistical heterogeneity in Federated Learning (FL) often leads to negative transfer, where a single global model fails to serve diverse client distributions. Personalized federated learning (pFL) aims to address this by tailoring models to individual clients. However, under most existing pFL approaches, clients integrate peer client contributions uniformly, which ignores the reality that not all peers are likely to be equally beneficial. Additionally, the potential for personalization at the instance level remains largely unexplored, even though the reliability of different peer models often varies across individual samples within the same client. We introduce FedDES (Federated Dynamic Ensemble Selection), a decentralized pFL framework that achieves instance-level personalization through dynamic ensemble selection. Central to our approach is a Graph Neural Network (GNN) meta-learner trained on a heterogeneous graph modeling interactions between data samples and candidate classifiers. For each test query, the GNN dynamically selects and weights peer client models, forming an ensemble of the most competent classifiers while effectively suppressing contributions from those that are irrelevant or potentially harmful for performance. Experiments on CIFAR-10 and real-world ICU healthcare data demonstrate that FedDES outperforms state-of-the-art pFL baselines in non-IID settings, offering robust protection against negative transfer.