FedFusion: Fusion Strategies in Federated Learning

• FedFusion is a collection of federated learning mechanisms that fuse diverse information sources at different pipeline stages to address challenges like non-IID data, multimodality, and label scarcity.
• The 2019 variant fuses frozen global and trainable local features to significantly reduce communication rounds, while the 2023 approach employs low-rank multimodal compression achieving 3–4× reduction in transmitted data.
• The 2025 framework enhances personalization through diversity- and cluster-aware encoder strategies coupled via similarity-weighted classifiers, improving accuracy, robustness, and fairness.

FedFusion denotes several distinct federated learning frameworks that share a common design intuition—explicitly combining information sources that standard federated averaging leaves weakly coupled—but instantiate that intuition in different ways. In the arXiv literature, the name has been used for at least three non-equivalent methods: a client-side feature-fusion extension of FedAvg for reducing communication rounds in non-IID federated learning, a manifold-driven multimodal framework for multi-satellite in-orbit fusion, and a transfer-learning framework with diversity- and cluster-aware encoders for heterogeneous, label-scarce clients (Yao et al., 2019, Li et al., 2023, Kahenga et al., 23 Sep 2025). The term therefore refers not to a single canonical algorithm, but to a family of method names centered on fusion at different levels of the federated pipeline.

1. Nomenclature and scope

The three uses of the name differ primarily in what is being fused, where the fusion occurs, and how communication efficiency is defined. In the 2019 formulation, fusion is local and architectural: each client fuses frozen global features with trainable local features before classification. In the 2023 formulation, fusion is multimodal and geometric: shallow features from HSI and LiDAR branches are compressed into a low-rank manifold-aware subspace for federated aggregation. In the 2025 formulation, fusion is interpersonal and adaptive: personalized encoders are retained locally while classifier heads are coupled through similarity-weighted aggregation and optional clustering (Yao et al., 2019, Li et al., 2023, Kahenga et al., 23 Sep 2025).

Work	Fusion locus	Reported outcome
2019 FedFusion	Local/global feature maps on each client	More than 60% fewer communication rounds
2023 FedFusion	Multimodal shallow features in low-rank subspace	Average accuracy 94.35%; communication compressed by a factor of 4
2025 FedFusion	Personalized encoders and similarity-weighted classifier heads	Improved accuracy, robustness, and fairness with comparable budgets

This multiplicity matters methodologically. A reference to “FedFusion” is ambiguous unless the surrounding problem class is specified: generic non-IID FL, multimodal remote sensing, or heterogeneous-feature transfer learning. This suggests that the label has functioned as a reusable motif rather than a standardized protocol.

2. Client-side feature fusion in federated averaging

The 2019 FedFusion was introduced as one of two mechanisms intended to reduce the communication cost and performance degradation of FedAvg, especially under non-IID client data (Yao et al., 2019). In vanilla FedAvg, the server broadcasts a single global model $G_r$, each selected client initializes its local model from $G_r$, performs $E$ local epochs of SGD, and returns the updated model for weighted averaging. FedFusion changes the client-side training architecture by preserving a frozen global feature extractor $E_g^G$ on the client while simultaneously training a local feature extractor $E_l^L$. The two feature maps are fused by a learnable module $F$, and a shared classifier $C$ operates on the fused representation. Only $(E_l^L,F,C)$ are updated locally; $E_g^G$ remains frozen.

Formally, for input $x$, the local and global feature maps satisfy

$G_r$0

Three differentiable fusion operators are defined. The $G_r$1 convolutional operator is

$G_r$2

with learnable $G_r$3. The per-channel weighted sum is

$G_r$4

where $G_r$5 is broadcast across spatial dimensions. The scalar weighted sum is

$G_r$6

with $G_r$7. In all cases, the fusion output is classified by $G_r$8, and gradients are propagated only through $G_r$9, $E$0, and $E$1.

The operational distinction from related baselines is explicit. FedAvg discards the global model after initialization. FedMMD retains both local and global subnetworks and constrains them through

$E$2

but does not fuse intermediate features. FedFusion instead blends local and global representations at the feature level on every batch. In the pseudocode reported for FedFusion, the client computes

$E$3

and backpropagates only $E$4.

The communication analysis is notable because per-round payload remains of order $E$5, as in FedAvg and FedMMD, with only the small fusion module added to the model size. The paper therefore measures “communication cost” by the number of rounds required to reach a target accuracy rather than by raw bytes per round. Empirically, FedFusion achieved up to roughly $E$6–$E$7 fewer rounds than FedAvg, whereas FedMMD saved roughly $E$8–$E$9 in non-IID settings. On Permuted MNIST, the convolutional fusion operator required $E_g^G$0 rounds to reach $E_g^G$1 accuracy and $E_g^G$2 rounds to reach $E_g^G$3, compared with $E_g^G$4 and $E_g^G$5 for FedAvg. Final convergence accuracies were also competitive or superior across the reported partitions: for example, on artificial non-IID CIFAR10 partition (a), FedAvg reached $E_g^G$6, FedFusion+Single $E_g^G$7, FedFusion+Multi $E_g^G$8, and FedFusion+Conv $E_g^G$9; on IID CIFAR10, the corresponding values were $E_l^L$0, $E_l^L$1, $E_l^L$2, and $E_l^L$3.

A further reported property is adaptation of newly incoming clients. Because each client retains the frozen global extractor and learns only a local extractor plus fusion adapter, new clients can immediately exploit global features and converge more quickly on unseen data. The number of local epochs needed for a new client to reach convergence accuracy was approximately $E_l^L$4–$E_l^L$5 for FedAvg, about $E_l^L$6 for FedFusion+Single, about $E_l^L$7 for FedFusion+Multi, and about $E_l^L$8 for FedFusion+Conv. The computational overhead was limited to an additional forward pass through the frozen global extractor and a small backward pass on $E_l^L$9, with overall client-side FLOPs increasing by a few percent. The paper explicitly does not provide formal convergence theorems or proofs.

3. Manifold-driven multimodal fusion for multi-satellite federated learning

The 2023 FedFusion addresses a different setting: in-orbit multi-satellite, multi-modality fusion under resource constraints, privacy constraints, and non-IID data across sensing platforms (Li et al., 2023). The motivating examples are HSI and LiDAR sensors on different low-orbit satellites. Here the goal is to train a single global classifier without moving raw imagery off satellite. The framework combines shallow-feature manifold estimation, low-rank compression, multimodal fusion, and a federated averaging module designed for manifold data in a deep latent space.

Each client runs two parallel 2D-CNN branches on its local HSI or LiDAR input $F$0. At layer $F$1, the shallow feature map is

$F$2

followed by BatchNorm and ReLU. To estimate a common low-dimensional manifold, the client randomly samples spatial positions $F$3 and forms

$F$4

where $F$5 is the number of channels. These local samples contribute to a global covariance or scatter matrix

$F$6

whose eigendecomposition

$F$7

yields the principal subspace. The paper notes a Laplacian-eigenmap style interpretation,

$F$8

but states that the implementation uses the global covariance $F$9 and its top $C$0 eigenvectors.

Compression is then imposed on deeper feature maps. A local feature tensor $C$1 is approximated by a low-rank factorization

$C$2

with $C$3, $C$4, and $C$5. The multimodal fusion step uses both cascading and additive operations:

$C$6

and the final cross-modal combination is

$C$7

The compressed factors $C$8, or a subset of them, are transmitted rather than the full feature maps. The summary equates this to a nuclear-norm minimization

$C$9

After compression, each client attaches a local classification head $(E_l^L,F,C)$0 and optimizes a loss combining cross-entropy on the fusion branch, branch-consistency MSE, and $(E_l^L,F,C)$1 regularization:

$(E_l^L,F,C)$2

Federated optimization then proceeds via standard FedAvg:

$(E_l^L,F,C)$3

The reported efficiency gains are expressed in both model traffic and runtime. Replacing a full $(E_l^L,F,C)$4 feature map, approximately $(E_l^L,F,C)$5 MB in Float16 per round, with low-rank SVD factors yields a communication reduction of roughly

$(E_l^L,F,C)$6

with $(E_l^L,F,C)$7 reported as the best accuracy-versus-cost trade-off. On a Jetson TX2 constellation with $(E_l^L,F,C)$8 HSI and $(E_l^L,F,C)$9 LiDAR satellites, model-update traffic per client fell from $E_g^G$0 MB to $E_g^G$1 MB, a $E_g^G$2 reduction. On the same hardware, training time for a full FedFusion round fell from $E_g^G$3 s to $E_g^G$4 s when federated, approximately a $E_g^G$5 speedup, together with a net end-to-end runtime saving of $E_g^G$6, approximately $E_g^G$7 minutes.

The reported predictive results are equally specific. Across Houston2013, Trento, and MUUFL, the method achieved an average overall accuracy of $E_g^G$8. On the Augsburg SAR+HSI benchmark using Sentinel-1 and Sentinel-2 data, it achieved $E_g^G$9 OA and $x$0 AA. The paper states that these results surpass traditional SVM/CNN/RNN baselines by more than $x$1–$x$2 points, recent deep fusion methods such as Cross and CALC by $x$3–$x$4 points, and transformer-based multimodal fusion, MFT, by $x$5–$x$6 points in OA/AA/Kappa.

4. Diversity- and cluster-aware encoders under heterogeneous features and label scarcity

The 2025 FedFusion addresses a third regime: heterogeneous feature spaces, severe non-IID data, and scarce labels across clients (Kahenga et al., 23 Sep 2025). The motivating examples include hospitals that record different subsets of clinical variables and imaging sites with different devices or demographics. Standard FL assumes a shared feature space $x$7 across clients; the framework instead allows each client to observe a subset $x$8. The method unifies three mechanisms: diversity- and cluster-aware personalization, a two-step domain-adaptation and frugal-labeling pipeline, and similarity-weighted classifier coupling.

The framework has three actors: a central server, labelled teacher clients, and learner clients with partial or no labels. Each communication round has two phases. In encoder pretraining, teacher clients train encoder-plus-task module $x$9 on labels, while unlabelled clients train $G_r$00 on a self-supervised pretext task such as rotation prediction. Encoders $G_r$01 are uploaded and aggregated by size-weighted averaging,

$G_r$02

and the resulting global encoder is broadcast. In task fine-tuning, each client attaches a local classifier head $G_r$03, optimizes supervised loss $G_r$04 on labelled data and pseudo-label loss $G_r$05 on unlabelled data when confidence exceeds $G_r$06, and returns its head for server-side coupling.

The personalization mechanism is instantiated in three variants. In DivEn, client $G_r$07 uses model $G_r$08, latent representation $G_r$09, and head parameters $G_r$10, optimizing

$G_r$11

with

$G_r$12

The server computes pairwise similarities

$G_r$13

and builds a client-specific aggregated head

$G_r$14

DivEn-mix uses the same loss but resets each client head after aggregation,

$G_r$15

DivEn-c adds clustering over feature subsets by Jaccard similarity, refining clusters until intra-cluster overlap is at least $G_r$16, followed by intra-cluster parameter averaging and inter-cluster similarity-weighted coupling. The corresponding cluster coherence term is

$G_r$17

The frugal-labeling pipeline is explicit. In step 1, labelled clients minimize

$G_r$18

while unlabelled clients minimize a rotation-based pretext objective

$G_r$19

In step 2, with weak and strong augmentations $G_r$20, the supervised and pseudo-label terms are

$G_r$21

$G_r$22

where $G_r$23. Fully labelled clients use $G_r$24, partially labelled clients use $G_r$25, and fully unlabelled clients use $G_r$26 while updating only $G_r$27 to limit encoder drift. An optional negative-transfer guard rolls a client back to a saved checkpoint if local accuracy drops below the previous round’s baseline.

The communication and computation profile differs from the earlier FedFusion variants. Encoder aggregation remains the same order as FedAvg, and classifier coupling requires sending only head parameters, which are small relative to encoders. The extra communication consists of latent summaries $G_r$28 for similarity computation, while the server incurs an $G_r$29 similarity-matrix step and one-time clustering overhead of $G_r$30. The paper characterizes the method as adding one extra vector exchange per round, preserving the same order of local compute, and keeping local computation within $G_r$31 of FedAvg.

The reported empirical picture is broad. On feature-heterogeneous tabular data, DivEn and DivEn-mix improved over Single, Auto_enc, Align_corr, and Class_agg. For Obesity with $G_r$32 features, the accuracies were $G_r$33 for Single, $G_r$34 for Auto_enc, $G_r$35 for Align_corr, $G_r$36 for Class_agg, $G_r$37 for DivEn, and $G_r$38 for DivEn-mix; with $G_r$39 features the corresponding values were $G_r$40, $G_r$41, $G_r$42, $G_r$43, $G_r$44, and $G_r$45. In the cluster-aware setting with $G_r$46 features, DivEn-c reached $G_r$47 on Obesity and $G_r$48 on Heart, compared with $G_r$49 and $G_r$50 for Single. The paper states that DivEn-c scales best as feature overlap increases and boosts minority clients by up to $G_r$51 over Single. For label-scarce and imaging-domain adaptation settings, it further reports that confidence-filtered pseudo-labels yield $G_r$52–$G_r$53 over purely adversarial or voting-based adaptation, and that FedFusion reduces the client-level standard deviation of accuracy by $G_r$54–$G_r$55.

5. Comparative interpretation of the three frameworks

Despite the shared name, the three frameworks intervene at different layers of the FL stack. The 2019 method fuses representations inside a client’s model while keeping server aggregation close to FedAvg. The 2023 method fuses modalities and compresses feature tensors before federated exchange. The 2025 method preserves personalized encoders and fuses information through similarity-weighted coupling of classifier heads and, optionally, cluster structure (Yao et al., 2019, Li et al., 2023, Kahenga et al., 23 Sep 2025).

Variant	Primary object of fusion	Communication interpretation
2019	Frozen global and trainable local feature maps	Same per-round order as FedAvg; fewer rounds to target accuracy
2023	HSI/LiDAR shallow features and low-rank factors	Explicit 3–4× compression of transmitted representation
2025	Personalized encoder outputs and classifier heads	Small extra exchange for heads and latent summaries

These contrasts matter because “communication efficiency” is not defined identically across the three lines of work. In the 2019 formulation, the per-round payload remains essentially unchanged and the savings appear as reduced rounds to reach a target accuracy. In the 2023 formulation, communication is reduced directly by transmitting low-rank factors instead of full feature tensors. In the 2025 formulation, efficiency is preserved by restricting extra exchange to heads and latent summaries while retaining encoder aggregation of the usual FedAvg order. A plausible implication is that the label “FedFusion” consistently signals a corrective mechanism for heterogeneity, but not a unique systems profile.

Another important distinction concerns personalization. The 2019 method personalizes only indirectly through a local extractor that is always fused with a frozen global one. The 2023 method personalizes through local multimodal feature extraction and compression while training toward a common classifier. The 2025 method is explicitly personalized: clients maintain encoders tailored to their local feature subsets and only the head structure is coupled through similarity and clustering. Consequently, the three methods speak to different failure modes of vanilla FL: non-IID optimization instability, bandwidth-constrained multimodal sensing, and joint feature-space mismatch plus label scarcity.

6. Limitations, unresolved questions, and historiographic significance

The limitations differ as sharply as the methods themselves. The 2019 FedFusion does not include formal convergence guarantees; the paper states that no theorem or proof is supplied and supports the method empirically instead (Yao et al., 2019). The 2025 framework identifies two explicit technical constraints: similarity computation is $G_r$56 on the server, albeit mitigated by clustering, and pseudo-label bias depends on the confidence threshold $G_r$57, for which a schedule is recommended (Kahenga et al., 23 Sep 2025). It also lists future directions including asynchronous updates, dynamic client joins and leaves, explainable local encoders, and benchmarks for mixed feature/label heterogeneity. The 2023 framework reports strong empirical compression and runtime gains but rests on a low-rank manifold assumption; this suggests that its effectiveness is tied to how well multimodal shallow features admit compact subspace structure (Li et al., 2023).

Historically, the three FedFusion papers illustrate an expansion of the fusion concept within federated learning. The 2019 work treats fusion as an internal architectural device for reconciling global and local representations under non-IID drift. The 2023 work reinterprets fusion through multimodal remote-sensing geometry and communication-aware low-rank modeling. The 2025 work extends the term again, using it for an overview of personalization, domain adaptation, frugal labeling, and classifier coupling. Read together, these works show that “FedFusion” in the literature names a sequence of distinct strategies for preserving local specificity while recovering shared structure in decentralized learning.