One-shot and few-round federated learning (FL) reduce communication burden by constructing a global model within one or a few communication rounds, but often achieve lower performance than conventional multi-round FL. Without frequent synchronization, client models may drift apart, causing neurons that learn similar features to occupy different positions across clients. At the same time, local data heterogeneity can cause corresponding parameters to differ in their importance to each client's predictions. Together, these challenges cause direct averaging to combine mismatched neurons and assign equal weight to parameters despite differences in their importance. We propose FedFiOT, a Fisher-weighted optimal transport framework for one-shot and few-round FL that leverages Fisher information to quantify parameter importance and guide neuron alignment and model aggregation. FedFiOT jointly optimizes neuron correspondences and global model parameters through alternating minimization of a shared Fisher-weighted objective. Theoretically, we show that FedFiOT's optimization procedure monotonically decreases the layer-wise merging objective and establish a stationarity bound that quantifies the effects of client heterogeneity and nonuniform Fisher weighting.Experiments across multiple datasets demonstrate consistent improvements over competing baselines in both one-shot and few-round federated learning settings.
Figures & tables
Figure 1: Overview of FedFiOT. At each layer, the server minimizes the joint objective Jℓ by alternating between Fisher-weighted neuron alignment and Fisher-weighted model merging. The merged layer serves as the alignment reference for the next iteration. This repeats for up to S iterations, or until no client’s alignment changes. Here, ℓ indexes the layer and [s] the iteration; wiℓ and wk,jℓ denote the weights of global neuron i and of neuron j of client k . In each permutation matrix, rows are global neurons, columns are client neurons, and orange cells mark matched pairs. wk,iℓ and fk,iℓ,λ denote the weights and damped Fisher block of the neuron of client k matched to global neuron i . For illustration, we provide a simple numerical example in Appendix G .
Dataset
α
FedAvg
OTFusion
WB Fusion
PFNM
RegMean
DENSE
Fisher Merge
FedFisher (K-FAC)
FedFiOT (Ours)
SVHN
0.2
54.4 ± 2.8
54.8 ± 0.9
60.8 ± 1.0
69.5 ± 1.3
61.6 ± 4.3
70.6 ± 0.9
50.7 ± 1.3
72.2 ± 1.8
78.9 ± 1.9
0.1
24.3 ± 2.8
31.0 ± 1.5
34.7 ± 2.3
44.1 ± 2.8
41.1 ± 5.0
48.4 ± 1.1
38.4 ± 3.2
61.6 ± 3.6
66.8 ± 0.8
CIFAR10
0.2
34.0 ± 0.8
37.0 ± 1.2
40.3 ± 1.2
46.0 ± 2.3
39.2 ± 1.7
42.5 ± 1.7
31.9 ± 6.7
48.7 ± 1.1
52.0 ± 0.9
0.1
30.3 ± 1.0
32.3 ± 2.8
29.0 ± 1.1
32.8 ± 1.3
30.9 ± 2.5
33.3 ± 1.8
25.2 ± 2.8
38.6 ± 2.1
43.3 ± 1.8
CINIC10
0.2
31.9 ± 1.1
34.4 ± 1.3
36.4 ± 1.1
39.6 ± 1.4
31.6 ± 1.5
33.2 ± 1.2
26.4 ± 4.6
40.5 ± 2.9
45.6 ± 1.4
0.1
26.6 ± 1.2
29.2 ± 0.7
25.4 ± 0.8
29.8 ± 1.2
25.6 ± 1.3
26.0 ± 0.4
25.6 ± 0.1
32.0 ± 0.5
39.6 ± 0.3
Table 1: One-shot test accuracy (%) with the number of clients K=6 fixed and the heterogeneity parameter α varied. Mean ± std over three seeds. FedFiOT uploads quantized weights and diagonal Fishers.
Figure 2: Test accuracy (%) over four rounds of FL with K=6 clients and α=0.1 .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
α
32-bit
16-bit
Difference
SVHN
0.2
78.2 ± 2.0
78.9 ± 1.9
+0.7
0.1
62.1 ± 2.1
66.8 ± 0.8
+4.7
CIFAR-10
0.2
51.1 ± 0.9
52.0 ± 0.9
+0.9
0.1
44.6 ± 1.3
43.3 ± 1.8
−1.3
CINIC-10
0.2
45.5 ± 2.5
45.6 ± 1.4
+0.1
0.1
38.6 ± 0.4
39.6 ± 0.3
+1.0
Appendix
Table 2: One-shot test accuracy (%) of FedFiOT with full-precision (32-bit) and quantized (16-bit) uploads, K=6 clients. Mean ± std over three seeds.
Method
Accuracy
FedAvg
25.5 ± 1.7
OTFusion
27.8 ± 2.1
PFNM
20.5 ± 3.0
RegMean
26.6 ± 2.6
DENSE
26.4 ± 0.7
FisherMerge
28.0 ± 3.0
Appendix
Table 3: One-shot test accuracy (%) on CIFAR-10 with K=6 clients and α=0.05 . Mean ± std over three seeds.
Figure 3: One-shot test accuracy (%) of FedFiOT and FedBEns with six clients, mean ± sample standard deviation over three training seeds. FedFiOT returns a single model; each FedBEns variant returns a five-model ensemble. FedBEns Kron selects server checkpoints on validation data.
FedFiOT
FedBEns
Models trained per client
1
5
Upload per client
1× FedAvg
>5× FedAvg
Models stored at inference
1
5
Forward passes per prediction
1
5
Appendix
Table 4: Cost comparison between FedFiOT and FedBEns per client and at inference, relative to FedAvg.
SVHN
CIFAR-10
CINIC-10
Method
α=0.2
α=0.1
α=0.2
α=0.1
α=0.2
α=0.1
FedAvg
41.2 ± 4.3
17.6 ± 1.3
16.3 ± 2.5
18.9 ± 2.8
14.9 ± 0.8
19.6 ± 2.6
OTFusion
55.6 ± 0.5
30.2 ± 1.4
36.5 ± 1.0
30.6 ± 0.4
33.8 ± 0.5
28.4 ± 1.3
WB Fusion
59.7 ± 2.6
34.3 ± 0.3
38.6 ± 1.6
27.4 ± 1.7
35.7 ± 0.9
25.8 ± 1.1
PFNM
68.9 ± 3.1
39.3 ± 0.3
42.8 ± 2.7
33.8 ± 1.4
38.4 ± 1.5
25.4 ± 0.9
RegMean
32.0 ± 1.6
25.4 ± 2.6
20.8 ± 1.4
18.2 ± 1.1
16.7 ± 1.1
17.8 ± 1.1
Appendix
Table 5: One-shot test accuracy (%) with independently initialized clients, K=6 . Mean ± std over three seeds.
Figure 4: Fraction of hidden neurons (%) permuted by FedFiOT in each round, K=6 clients and α=0.1 . Lines show the mean over three seeds and shaded regions show ± one standard deviation.
SVHN
CIFAR-10
CINIC-10
Variant
α=0.2
α=0.1
α=0.2
α=0.1
α=0.05
α=0.2
α=0.1
(A) FedFiOT
78.9 ± 1.9
66.8 ± 0.8
52.0 ± 0.9
43.3 ± 1.8
41.4 ± 1.7
45.6 ± 1.4
39.6 ± 0.3
(B) No alignment
60.4 ± 2.7
42.4 ± 5.6
40.1 ± 4.2
33.2 ± 1.5
27.6 ± 3.9
33.1 ± 2.0
28.2 ± 1.6
(C) Euclidean matching
77.3 ± 0.2
49.9 ± 4.7
50.0 ± 1.3
34.4 ± 1.8
27.7 ± 4.1
42.2 ± 1.2
32.0 ± 0.5
(D) Uniform aggregation
61.3 ± 1.7
38.1 ± 1.0
38.2 ± 1.9
31.7 ± 1.2
21.1 ± 0.7
34.4 ± 1.0
26.2 ± 0.9
(E) WB Fusion
60.8 ± 1.0
34.7 ± 2.3
40.3 ± 1.2
29.0 ± 1.1
18.3 ± 0.7
36.4 ± 1.1
25.4 ± 0.8
Appendix
Table 6: Ablation of FedFiOT’s components on K=6 . Test accuracy (%), mean ± std over three seeds.
Merge
Alignment
w⊤
J
Average by index
identity
(5,3,7)
108
Fisher merging
identity
(2.5,3,7)
83
Average of aligned weights
(2,3,1)
(1,5,9)
12
FedFiOT
(2,3,1)
(0.5,5.5,8.5)
9
Appendix
Table 7: Merged layer and objective J for the three-neuron example. Methods without alignment use identity alignments for both clients.
Federated learning aggregation combines client updates into shared or personalized models while balancing data heterogeneity, robustness, privacy, and communication.
Heterogeneous federated learning requires clients with diverse computational capacities to collaboratively train a global model, where each client trains a capacity-constrained submodel. Existing methods select submodel parameters using heuristic importance measures---most prominently parameter magnitude---without theoretical justification for why these measures support convergence. We identify a fundamental gap: existing parameter selection criteria lack theoretical grounding in the convergence framework, partial client participation introduces additional estimation effects in the Fisher scores. We propose FedFIbOS: Fisher Importance-based Optimal Submodelling for heterogeneous federated learning, using Fisher Information in a principled criterion derived from minimizing submodel masking error. %We formally establish when magnitude selection is equivalent to Fisher selection fail under non-IID heterogeneous federated learning. We theoretically formulate submodel selection through a Fisher-weighted quadratic masking surrogate and show that the raw Fisher top-k rule implemented by FedFIbOS solves this surrogate under a Fisher-dominant ranking condition. The resulting method retains the convergence structure of the underlying masked federated optimization bound. Fisher scores are efficiently estimated from empirical diagonal Fisher information using squared gradients, enabling stable and adaptive parameter selection without additional optimization overhead. Experiments on CIFAR-10, CIFAR-100, and AGNews under pathological and Dirichlet non-IID settings show FedFIbOS achieves ≈10% higher accuracy than the state of the art, with improvements becoming more pronounced under stronger heterogeneity.
Yasmeen Afzal, Jeremiah D. Deng, Haibo Zhang
School of Computing, University of Otago, New Zealand · School of Computer Science and Engineering, University of New South Wales, Australia
One-Shot Federated Learning (OSFL) addresses extreme communication regimes in which clients interact with the server only once, amplifying the impact of heterogeneous client data distributions. In particular, the interaction of domain shift and label shift across clients induces misaligned feature representations that cannot be corrected through iterative optimization. Existing OSFL methods rely on distillation, server-side generation or ensemble-based aggregation, but assume aligned representations or address domain and label shift separately. We introduce SLOT-Align (Single-round, Learning-free Optimal Transport Alignment), a geometry-aware feature harmonization framework for OSFL. SLOT-Align uses a shared frozen encoder to extract compact feature statistics, constructs a global reference via Bures-Wasserstein barycenters, and aligns local representations using closed-form geodesic optimal transport maps. The method is computationally efficient and can be combined with existing OSFL pipelines relying on frozen encoders without modifying their training procedures. Extensive experiments across multiple benchmarks, pretrained backbones, and OSFL methods show that SLOT-Align consistently improves accuracy and robustness under joint domain and label shift.
Daniele Berardini, Vito Paolo Pastore, Vittorio Murino
AI for Good · AI for Good (AIGO), Italian Institute of Technology, Genoa, Italy · MaLGa-DIBRIS, University of Genoa, Genoa, Italy +1
Federated learning usually aggregates client updates using dataset size or gradient-level criteria, while overlooking internal signals about how each client model is organizing its input space during training. We introduce FedOUI, a simple aggregation rule based on the Overfitting-Underfitting Indicator (OUI), an activation-based and label-free metric. Each participating client sends its local update together with a OUI value computed on a fixed probe batch, and the server estimates the round-wise OUI distribution to assign lower weights to structurally atypical clients through a smooth reweighting rule. We evaluate FedOUI on CIFAR-10 under strong non-IID partitioning and noisy-client conditions, comparing it with FedAvg, FedProx, and a gradient-alignment baseline. The clearest gains appear under strong heterogeneity, where OUI-based weighting improves aggregation quality while remaining lightweight and interpretable. These results show that internal activation structure can provide useful information for federated aggregation beyond client size and gradient geometry.
Alberto Fernández-Hernández, Jose I. Mestre, Cristian Pérez-Corral +3
Universitat Politècnica de València, Valencia, Spain · Universitat Jaume I, Castelló de la Plana, Spain · Openchip & Software Technologies S.L., Spain