---
title: 'FedDCL: Federated Data Collaboration and Continual Learning'
url: https://www.emergentmind.com/topics/feddcl
type: topic
---

# FedDCL: Federated Data Collaboration and Continual Learning

Searching arXiv for “FedDCL” and related variants to ground the article in current papers.
FedDCL is an acronym used in arXiv literature for more than one federated-learning framework. In its 2024 formulation, it denotes **federated data collaboration learning**, a hybrid privacy-preserving framework that combines **data collaboration analysis** with **federated learning** so that user institutions share dimensionality-reduced intermediate representations with intra-group servers, and only those servers participate in iterative federated optimization [2409.18356]. In a later 2025 formulation, the same acronym denotes a **data-free continual learning** framework for **model-heterogeneous federated learning**, built around diffusion-derived class prototypes, generative replay, and collaborative distillation [2509.25977]. The term therefore has a stable federated-learning context but not a single universally fixed technical meaning.

## 1. Nomenclature and scope

In the available literature here, “FedDCL” most explicitly refers to the framework titled **“FedDCL: a federated data collaboration learning as a hybrid-type privacy-preserving framework based on federated learning and data collaboration”** [2409.18356]. That paper defines FedDCL as a method for **privacy-preserving integrated analysis across multiple institutions** when standard federated learning is difficult to deploy because user institutions cannot maintain **continuous communication with an outside server**.

The acronym is later reused by **“Data-Free Continual Learning of Server Models in Model-Heterogeneous Federated learning”**, which also uses the name FedDCL but targets a different setting: **class-incremental continual learning**, **model heterogeneity**, **catastrophic forgetting**, and **knowledge misalignment** [2509.25977]. Nearby acronym variants add further ambiguity. **DCFL** is used for **non-IID awareness Data Condensation aided Federated Learning** [2312.14219] and for **Decoupled Contrastive Learning for Federated Learning** [2508.04005]. A related but differently named line is **Federated Deep Subspace Clustering**, which decentralizes deep clustering by sharing only an encoder while keeping clustering-specific components local [2501.00230].

| arXiv id | Expansion | Primary technical focus |
|---|---|---|
| [2409.18356] | federated data collaboration learning | communication-constrained privacy-preserving integrated analysis |
| [2509.25977] | data-free continual learning of server models | continual learning under model heterogeneity |
| [2312.14219] | Non-IID awareness Data Condensation aided Federated Learning | client grouping and condensed-data exchange |
| [2508.04005] | Decoupled Contrastive Learning for Federated Learning | alignment/uniformity decoupling in contrastive FL |

A common source of confusion is therefore terminological rather than methodological. FedDCL is not a single canonical optimizer, loss, or architecture; it is an acronym attached to distinct federated frameworks with different assumptions and objectives.

## 2. Federated data collaboration learning as a hybrid FL–DC architecture

The 2024 FedDCL framework is motivated by multi-institutional settings, especially those in which **raw data cannot be shared because of privacy and cost concerns**, some institutions may be on **isolated networks**, and analysis on a single institution is often too small to achieve good performance, especially for **rare diseases** [2409.18356]. Standard federated learning methods such as **FedSGD** and **FedAvg** preserve raw-data privacy but require **repeated communication rounds**, which the paper identifies as a major deployment obstacle when continuous external communication is extremely difficult.

FedDCL addresses this by combining **Data Collaboration Analysis (DC)** with **Federated Learning (FL)**. The framework has three roles. **User institutions** hold local datasets and private mapping functions. **Intra-group DC servers** receive intermediate representations from institutions in the same group, perform representation alignment, and communicate externally. A **central FL server** coordinates federated learning across the group servers. The critical architectural point is that **the user institutions do not need external iterative communication; only the intra-group servers do**.

The user-level data are written as
$$
X_j^{(i)} \in \mathbb{R}^{n_{ij}\times m}, \quad Y_j^{(i)} \in \mathbb{R}^{n_{ij}\times \ell},
$$
where \(i=1,\dots,d\) indexes groups and \(j=1,\dots,c_i\) indexes institutions within group \(i\). Each institution computes a private mapping \(f_j^{(i)}\), transforming raw data into intermediate representations
$$
\widetilde{X}_j^{(i)} = f_j^{(i)}(X_j^{(i)}),
$$
and similarly transforms a shared anchor dataset \(A\) into
$$
\widetilde{A}_j^{(i)} = f_j^{(i)}(A).
$$

The collaboration layer is then constructed inside each group. Institutions send \(\widetilde{X}_j^{(i)}\), \(\widetilde{A}_j^{(i)}\), and \(Y_j^{(i)}\) to their intra-group DC server. The DC server converts these into **collaboration representations** that are compatible across institutions and groups. Federated learning is then run over those collaboration representations rather than over raw data.

This makes FedDCL a hybrid design rather than a minor variation of FedAvg. The DC component solves a representation-alignment problem, while the FL component solves the group-level model-training problem.

## 3. Mathematical construction and training workflow

FedDCL proceeds in five explicit steps [2409.18356]. First, all users generate the same shareable pseudo anchor dataset
$$
A \in \mathbb{R}^{r\times m},
$$
where \(r\) is the number of anchor samples. The paper describes this as public or dummy data, randomly constructed, while noting that better performance may come from anchor data closer to the raw distribution.

Second, each institution constructs dimensionality-reduced intermediate representations
$$
\widetilde{X}_j^{(i)} = f_j^{(i)}(X_j^{(i)}), \qquad \widetilde{A}_j^{(i)} = f_j^{(i)}(A),
$$
with \(\widetilde{m}_{ij} < m\). The paper states that these mappings can be linear or nonlinear, including **PCA**, **LDA**, **t-SNE**, **Fisher discriminant variants**, and **locality-preserving methods**.

Third, each intra-group DC server constructs collaboration representations. Inside group \(i\), it concatenates anchor-side intermediate representations,
$$
\widetilde{A}^{(i)} = [\widetilde{A}^{(i)}_1,\dots,\widetilde{A}^{(i)}_{c_i}],
$$
computes a rank-\(\widehat{m}_i\) approximation by SVD,
$$
\widetilde{A}^{(i)} \approx U^{(i)}\Sigma^{(i)}(V^{(i)})^\top,
$$
and defines
$$
\widetilde{B}^{(i)} = U^{(i)} C_1^{(i)},
$$
where \(C_1^{(i)}\) is nonsingular. The central FL server then concatenates
$$
\widetilde{B} = [\widetilde{B}^{(1)},\dots,\widetilde{B}^{(d)}],
$$
computes another rank-\(\widehat{m}\) SVD,
$$
\widetilde{B} \approx P D Q^\top,
$$
and forms
$$
Z = P C_2,
$$
with nonsingular \(C_2\). This \(Z\) is returned to the intra-group servers.

Fourth, each DC server solves, for every institution,
$$
G_j^{(i)} = \arg\min_G \|\widetilde{A}_j^{(i)}G - Z\|_F,
$$
and constructs collaboration representations
$$
\widehat{X}^{(i)}_j = \widetilde{X}^{(i)}_j G_j^{(i)}.
$$
At group level,
$$
\widehat{X}^{(i)} =
\begin{bmatrix}
\widehat{X}^{(i)}_1\\
\vdots\\
\widehat{X}^{(i)}_{c_i}
\end{bmatrix},
$$
and the integrated learning target becomes
$$
h(\widehat{X}) \approx Y.
$$

Fifth, the final model returned to institution \((i,j)\) is
$$
t_j^{(i)}(X) = h(f_j^{(i)}(X)G_j^{(i)}).
$$
The paper uses **FedAvg** in experiments, while stating that the framework can be combined with other FL algorithms.

The paper also gives a conditional equivalence result. If all mappings are linear,
$$
f_j^{(i)}(X)=XF_j^{(i)},
$$
and satisfy the shared-range condition
$$
\mathcal{R}(F^{(1)}_1)=\cdots=\mathcal{R}(F^{(c)}_{d_c}),
$$
then there exists a common \(F\) such that
$$
\widehat{X}=XF,\qquad \mathcal{R}(F)=\mathcal{F}.
$$
This means that, under those assumptions, FedDCL is equivalent to federated learning on a common dimensionality-reduced feature space. This suggests that the DC stage is not merely a preprocessing heuristic but can define a common latent representation under explicit structural conditions.

## 4. Privacy model, communication pattern, and empirical behavior

The 2024 FedDCL paper states that raw data \(X_j^{(i)}\) are never shared, and the user-specific mapping functions \(f_j^{(i)}\) remain private [2409.18356]. What is shared are intermediate representations \(\widetilde{X}_j^{(i)}\), anchor representations \(\widetilde{A}_j^{(i)}\), labels \(Y_j^{(i)}\) to intra-group servers, group-level compressed representations \(\widetilde{B}^{(i)}\) to the FL server, and then learned alignment matrices \(G_j^{(i)}\) and the final model \(h\) back to users. The paper describes this as a **double privacy layer**: the mapping \(f_j^{(i)}\) is private, and dimensionality reduction provides **\(\varepsilon\)-DR privacy**.

Communication reduction is the framework’s defining systems claim. User institutions do not participate in every FL round; the iterative communication occurs only between intra-group DC servers and the central FL server. The authors note that each user institution requires only **two cross-institutional communications**. This is why the method is positioned as especially useful when institutions are grouped and only group servers can communicate continuously with the outside world.

The experimental protocol uses a **fully connected neural network** with **batch size = 32**. **Centralized**, **Local**, and **DC** are trained for **40 epochs**, whereas **FedAvg** and **FedDCL** use **4 epochs per round, 20 rounds, total 80 epochs**. For **DC** and **FedDCL**, the intermediate representations are built by **PCA with random orthogonal mapping**, and the **anchor dataset size** is \(r=2000\). The proof-of-concept experiment uses **BatterySmall**, a regression task on lithium-ion battery sensor data, with **4 user institutions in 2 groups**, **100 samples per institution**, and \(\widehat{m}=4\). A broader evaluation uses **BatterySmall**, **CreditRating_Historical**, **eICU**, **HumanActivity**, **MNIST**, and **Fashion-MNIST**, distributed across **20 user institutions**, **5 groups**, and **4 institutions per group**. The reported metrics are **RMSE** for regression datasets and **Accuracy** for classification datasets.

The reported findings are qualitative but consistent: **FedDCL is much better than Local**, **comparable to FedAvg and DC**, and on **BatterySmall** it shows **higher convergence than FedAvg**. On the six-dataset study, it achieves **high recognition performance comparable to FedAvg and DC**. In the **MNIST** group-scaling experiment, accuracy increases as the number of groups increases, and FedDCL remains competitive. A plausible implication is that the dimensionality-reduced collaboration space may regularize optimization, which aligns with the authors’ suggestion that fewer effective parameters may explain the higher convergence observed on BatterySmall.

## 5. Data-free continual-learning FedDCL in model-heterogeneous FL

A later paper reuses the acronym for a different framework: **FedDCL** as **data-free continual learning of server models in model-heterogeneous federated learning** [2509.25977]. Here the setting is one server and \(K\) clients, with client-private datasets \(\mathcal{D}_k^{(t)}\) arriving over a **class-incremental** task stream \(\{t\}_{t=1}^T\), disjoint classes across tasks, **non-IID data** induced by \(Dir(\gamma)\), and **model heterogeneity**, so parameter averaging is not applicable. The server model \(\mathcal{M}_s\) is trained by distillation rather than weight aggregation.

The framework’s central mechanism is a **pre-trained diffusion model** with frozen noise predictor, used to learn lightweight class-specific prototypes \(p\). The standard diffusion forward process is written as
$$
q\left(x_{1:T} \mid x_0\right) = \prod_{t=1}^T q\left(x_t \mid x_{t-1}\right),
$$
with
$$
q\left(x_t \mid x_{t-1}\right)=\mathcal{N}\left(x_t;\sqrt{\alpha_t}x_{t-1},(1-\alpha_t)\mathbf{I}\right),
$$
and the reverse process is conditioned on class information \(\mathcal{C}\). Instead of optimizing the diffusion network, the method optimizes only the prototype through
$$
\mathcal{L}_{\text{proto}}(p) = \mathbb{E}_{t,\, x_0\sim p_{\text{data}},\, \epsilon\sim \mathcal{N}(\mathbf{0},\mathbf{I})} \left\| \epsilon-\epsilon_\theta(x_t,t,p) \right\|_2^2.
$$
Client-side class prototypes \(p_{i,k}\) are uploaded and averaged at the server as
$$
p_i = \frac{1}{K}\sum_{k=1}^K p_{i,k}.
$$

The abstract identifies three **data-free advantages**, and the detailed pipeline implements them directly: **generation of synthetic data for the current task to augment training and counteract non-IID data distributions**, **exemplar-free generative replay for retaining knowledge from previous tasks**, and **data-free dynamic knowledge transfer from heterogeneous clients to the server**. The method therefore targets four simultaneous issues: **data heterogeneity**, **model heterogeneity**, **catastrophic forgetting**, and **knowledge misalignment**.

Its evaluation uses two benchmarks: **Grayscale**, formed from the first 10 classes of **MNIST**, **EMNIST**, and **Fashion-MNIST** for **30 classes total**, and **RGB**, formed from the first 30 classes of **CIFAR-100**. Each benchmark is split into **3 class-incremental tasks**. There are **5 clients** with heterogeneous capacities: one **Large**, two **Medium**, and two **Small**. For Grayscale, the L model and server use a **4-layer CNN**, while M and S use two different **2-layer CNNs**. For RGB, the client models are **ResNet-50**, **ResNet-32**, and **ResNet-18**, and the server uses a **pre-trained ViT**, with only the classification head fine-tuned. The metrics are **ACC** and the forgetting measure
$$
F_t=\frac{1}{t-1}\sum_{j=1}^{t-1} f_j^t,\quad f_j^t=\max_{l\in\{1,\dots,T-1\}} acc_{l,j}-acc_{t,j}.
$$

The reported results are explicitly strong. On **Grayscale, \(Dir(\gamma=0.3)\)**, FedDCL reaches **80.92%** final cumulative accuracy with forgetting **4.37**; on **Grayscale, \(Dir(\gamma=0.5)\)**, it reaches **82.99%** with forgetting **4.86**. On **RGB**, it achieves **74.10%** under one heterogeneity setting and **73.43%** under the other, with forgetting **8.40** and **6.85**, respectively. The paper reports that improvements over the strongest baselines are **24.37%** and **18.96%** in the RGB setting. This establishes a second, distinct meaning of FedDCL: not communication restructuring through collaboration representations, but data-free continual distillation through prototype-conditioned synthetic data.

## 6. Relations to adjacent federated clustering, condensation, and contrastive methods

FedDCL sits within a broader family of methods that respond to FL constraints by changing what is shared, where alignment occurs, or how local objectives are regularized. **Federated Deep Subspace Clustering** introduces **FDSC**, a federated version of deep subspace clustering for **unlabeled, decentralized, and privacy-sensitive data**, in which only the **encoder** is federated while the **decoder** and **self-expressive layer** remain private to each client [2501.00230]. The global encoder is updated by
$$
E_*^{(t+1)} = \sum_{i=1}^m p_i E_i^t,
$$
while each client keeps local clustering structure through a self-expression matrix and a **locality-preserving adjacency regularizer**. This is conceptually close to the 2024 FedDCL emphasis on sharing only a representation-bearing component and retaining client-specific structure locally, although the problem class is unsupervised clustering rather than integrated supervised analysis.

**DCFL: Non-IID awareness Data Condensation aided Federated Learning** approaches heterogeneity differently. It groups clients with **CKA**, exchanges **condensed datasets** \(\widetilde{\mathcal{D}_k}\) rather than raw data, and adds a **CKA-guided client selection strategy**, **filtering mechanisms**, and **data enhancement techniques** [2312.14219]. Standard FL shares model updates; DCFL additionally shares **small synthetic/condensed datasets** to enrich training with complementary distributions and reduce communication rounds. This is closer to the 2025 FedDCL variant, where synthetic data and knowledge transfer also replace direct raw-data access, although the continual-learning and diffusion-prototype machinery are specific to that later FedDCL.

A different nearby line is **Decoupled Contrastive Learning for Federated Learning**, also abbreviated **DCFL**, which argues that standard supervised contrastive learning is mismatched to FL because the asymptotic assumption of infinitely many negatives fails on clients with small local datasets [2508.04005]. Its core loss is
$$
\mathcal{L}_{\mathrm{DCFL}} = -\lambda_a \sum_{p\in P_i}\frac{\mathrm{sim}(z_i,p)}{\tau} + \lambda_u |P_i|\log\sum_{n\in N_i}\exp\left(\frac{\mathrm{sim}(z_i,n)}{\tau}\right),
$$
with \(\lambda_a+\lambda_u=1\). This explicitly decouples alignment and uniformity and serves as a client-side regularizer rather than a communication redesign.

Taken together, these methods show that “FedDCL” belongs to a wider design space of federated systems that modify one of three layers: **representation exchange**, **synthetic-data exchange**, or **local objective geometry**. The 2024 FedDCL primarily restructures communication through **intra-group DC servers** and **collaboration representations** [2409.18356]. The 2025 FedDCL primarily restructures continual knowledge transfer through **diffusion-derived prototypes** and **data-free distillation** [2509.25977]. The surrounding literature indicates that the core research problem is not a single algorithmic motif but the broader question of how federated learning should be redesigned when privacy, heterogeneity, communication scarcity, and task dynamics jointly invalidate the assumptions of plain parameter averaging.

Source: https://www.emergentmind.com/topics/feddcl