FedMate: Federated Learning for Heterogeneous Data

• FedMate is a federated learning method that addresses cross-client heterogeneity through a bilateral optimization framework combining server-side prototype recalibration and client-side merit-based discrimination.
• It enhances global consistency and local personalization by decoupling classifier and feature training while employing cost-aware feature transmission to reduce communication overhead.
• FedMate demonstrates superior performance in classification and semantic segmentation by harmonizing unbiased global aggregation with effective local adaptation across diverse datasets.

FedMate is a federated learning method designed for cross-client data heterogeneity, especially the non-i.i.d. and pathological splits that bias consensus formation and weaken the complementary fusion of generalization- and personalization-oriented knowledge. It addresses limitations attributed to model decoupling and representation center loss approaches by combining server-side prototype re-calibration and classifier refinement with client-side merit-discrimination training and cost-aware feature transmission. In the source paper, this design is described as a bilateral optimization framework that aims to harmonize global consistency, local adaptability, and communication efficiency across classification and semantic segmentation settings (Yang et al., 25 Aug 2025).

1. Problem setting and motivation

Federated learning in heterogeneous environments faces degraded generalization, slow convergence, and poor adaptation, particularly when clients exhibit disjoint or imbalanced class distributions. FedMate is motivated by the need to robustly mitigate statistical heterogeneity, enable a balance between generalization and personalization, and improve communication and computational efficiency (Yang et al., 25 Aug 2025).

The motivating diagnosis is that existing approaches often rely on static and restricted metrics to evaluate local knowledge and adopt global alignment too rigidly. In the formulation associated with FedMate, these choices can distort consensus and diminish model adaptability. The method therefore treats aggregation not as a purely uniform averaging problem, but as a re-calibration problem in which local knowledge is weighted according to multiple signals and then fused with explicit category structure.

A central premise is that heterogeneous clients contain both bias-inducing and complementary information. FedMate is built to preserve the latter while suppressing the former. This is reflected in two linked goals: preserving global consistency through recalibrated class-level structure, and retaining local specialization by avoiding unnecessary synchronization of the feature extractor.

2. Bilateral optimization framework

FedMate is organized around a server-side branch and a client-side branch, each with distinct responsibilities. The server constructs a dynamic global prototype and performs category-wise classifier fine-tuning. The clients carry out local prototype estimation, decoupled classifier and feature training, complementary classification fusion, merit-based discrimination training, and selective feature transmission (Yang et al., 25 Aug 2025).

Component	Purpose
Dynamic Prototype Aggregation	Weighted average of client prototypes for each class; lessens bias from class imbalance/disjoint sets
Category-wise Classifier Fine-tune	Global classifier refined per class using recalibrated prototypes
Complementary Classification Fusion (CCF)	Discriminator-based adversarial alignment of local/global outputs for de-biasing & anti-forgetting
Merit-based Discrimination	Prevents low-quality/biased updates from dominating
Cost-aware Feature Transmission (CFT)	Triggers expensive feature extractor uploads only when necessary to save bandwidth

This organization is explicitly bilateral rather than symmetric. The server is responsible for consensus condensation under heterogeneity, while the client is responsible for preserving adaptation under global guidance. A plausible implication is that FedMate treats global aggregation and local personalization as coupled but non-identical optimization problems.

The framework also decouples feature and classifier learning locally. According to the summary, the classifier is trained with the feature extractor frozen, and the feature extractor is then trained with the classifier frozen while being aligned to the global prototype. This decoupling is used to constrain drift without collapsing client-specific adaptation.

3. Server-side prototype re-calibration and classifier refinement

On the server side, FedMate constructs a category-wise global prototype in the shared feature space and updates it dynamically each round. Local prototypes $P_i^{t+1}$ from participating clients $\mathcal{Z}_t$ are aggregated as

$$P^{t+1} = \frac{\sum_{i\in \mathcal{Z}_t} \alpha_i P_i^{t+1}}{\sum_{i\in \mathcal{Z}_t} \alpha_i}.$$

The aggregation weights $\alpha_i$ are described as reflecting data share, class representation, or other factors. In the abstract, the calibration is further characterized as a holistic integration of sample size, current parameters, and future prediction (Yang et al., 25 Aug 2025).

This prototype serves as the basis for global consistency. Rather than relying only on model averaging, FedMate explicitly aggregates class centers, which is intended to reduce noise from outlier or under-represented categories and to improve compactness and separation of class clusters. In the paper’s interpretation, this is critical for unbiased consensus condensation under severe heterogeneity.

The global classifier is then aggregated and fine-tuned in a category-wise manner. The classifier aggregation is given by

$$\phi^{t+1} = \frac{\sum_{i\in \mathcal{Z}_t} \alpha_i \phi_i^{t+1}}{\sum_{i\in \mathcal{Z}_t} \alpha_i},$$

and the category-wise fine-tuning step is described as

$$\phi^{t+1}_k \gets \arg\min_{\phi_k} \ell_{\text{ce}}(\phi_k; P^{t+1}_k).$$

This procedure uses the recalibrated prototype to adapt category-specific parameters after aggregation. The source characterizes that step as preserving global consistency while improving robust adaptation.

Feature extractor aggregation is conditional rather than mandatory. When the cost-aware scheme is triggered, the server aggregates feature extractors from selected rounds as

$$\theta^{t+1} = \frac{\sum_{i\in \mathcal{Z}_t} \alpha_i \theta_i^{t+1}}{\sum_{i\in \mathcal{Z}_t} \alpha_i},$$

and otherwise retains the current model to maximize local adaptation. This conditionality distinguishes FedMate from approaches that enforce global synchronization every round.

4. Client-side local training, CCF, and merit-discrimination

Each client computes a per-class local prototype. For client $i$ and class $k$,

$$P_{i,k}^t = \frac{1}{|\mathcal{D}_{i,k}|} \sum_{x_{i,k}\in \mathcal{D}_{i,k}} f_i(\theta_i^t; x_{i,k}).$$

Local training proceeds in two stages. First, the classifier is trained with the feature extractor frozen:

$\mathcal{Z}_t$0

Second, the feature extractor is trained with the classifier frozen and, when a global prototype $\mathcal{Z}_t$1 is available, a centroid-alignment term is added:

$\mathcal{Z}_t$2

with total objective

$\mathcal{Z}_t$3

The parameter $\mathcal{Z}_t$4 is the weight on global prototype alignment, and the summary notes that over-tuning this can cause overfitting or underfitting.

Complementary Classification Fusion is the main client-side de-biasing mechanism. It introduces two discriminators: a Prototype Discriminator $\mathcal{Z}_t$5, which distinguishes real local versus global prototypes, and a Classification Discriminator $\mathcal{Z}_t$6, which tries to discern classifier output from local versus global (Yang et al., 25 Aug 2025). The adversarial process is described as follows: the discriminator tries to classify local as 0, while the generator, comprising classifier and feature components, aims for 1. This encourages local representations and decisions to become indistinguishable from global ones, thereby reducing class-level bias and catastrophic forgetting. The summary also states that local prototypes are used for cross-entropy loss to stabilize training.

Merit-based discrimination training is presented as a consequence of this adversarial alignment. By aligning local outputs through the discriminators, low-quality or biased prototypes and classifiers are discouraged from dominating aggregation. The method therefore does not merely transmit all client knowledge uniformly; it introduces a quality-sensitive filter at the level of representation and decision alignment.

5. Cost-aware feature transmission and efficiency

Cost-aware Feature Transmission is intended to balance model performance and communication efficiency. Instead of uploading feature extractors every round, clients do so only on a predefined schedule or when local adaptation diverges significantly; the parameter $\mathcal{Z}_t$7 controls the schedule (Yang et al., 25 Aug 2025).

The source attributes three effects to CFT. First, it reduces communication cost. Second, it preserves beneficial local adaptation by relaxing strict global generalization requirements. Third, on highly skewed local data, it can improve heterogeneous-test performance relative to transmitting feature extractors in every round. The paper summarizes this point by stating that “CFT achieves superior performance on the heterogeneous test set ... compared to transmitting feature extractors in every round (ALL)... by relaxing strict global generalization requirements...”.

Algorithmically, the server aggregates local classifiers, prototypes, and optionally features, then triggers fine-tuning and sends back updates. On the client, each local epoch freezes the feature extractor to train the classifier, then freezes the classifier to align the feature extractor with the global prototype if received, recomputes the local prototype, and selectively uploads according to CFT. This procedural structure links communication scheduling directly to representation control rather than treating communication as an external systems choice.

A common misconception would be to view CFT as a purely bandwidth-saving heuristic. The paper presents it as a mechanism that also preserves personalization by avoiding unnecessary feature-level synchronization. In that sense, the communication policy is part of the optimization design rather than a post hoc engineering modification.

6. Empirical behavior, scalability, and interpretation

FedMate is evaluated on five datasets of varying complexity. For generalization and adaptation, the summary explicitly lists CIFAR-10, CIFAR-100, CINIC-10, and EMNIST, and reports that the method substantially outperforms classical and personalized FL schemes across all heterogeneity levels, including pathological splits (Yang et al., 25 Aug 2025). The reported ablation study indicates that removing any one component degrades performance, covering CCF, feature aggregation, prototype supervisions, and cost-aware transmission. Convergence is described as stable and rapid, at approximately 150 rounds even on difficult datasets.

In system cost terms, the summary states that FedMate’s communication overhead is equivalent to model-only transmission, in contrast to more intensive prototype+model approaches. It also reports that the method achieves targeted accuracy with noticeably less wall-clock time than all baselines, attributing this to decoupling the extractor and classifier and aligning via global prototypes. This suggests that the architectural decomposition is intended to improve optimization efficiency as well as statistical robustness.

The semantic segmentation experiments are framed as evidence of real-world scalability. On Cityscapes for autonomous driving, using 152 clients, FedMate is reported to produce visually crisper segmentation boundaries and coherent class regions, outperforming strong FL baselines under high distribution skew. On PatternNet for remote sensing, it outperforms FedSeg, especially for fine-grained classes involving small, complex objects. The summary interprets these outcomes by emphasizing that the explicitly constructed, unbiased global prototype is central to segmentation performance because class-feature alignment is crucial for pixel-level tasks (Yang et al., 25 Aug 2025).

An objective reading of the paper places FedMate within the class of federated methods that seek simultaneous global generalization and local adaptation under heterogeneity. What distinguishes it in the source description is not only prototype use or model decoupling in isolation, but the combination of dynamic prototype aggregation, category-wise classifier fine-tuning, adversarial complementary fusion, merit-based discrimination, and selective feature transmission. A plausible implication is that FedMate reframes heterogeneous federated learning as a coordinated problem of representation re-calibration, class-aware aggregation, and communication-aware personalization rather than simple averaging under non-i.i.d. noise.