FedDAG: Diverse Federated Research Paradigms

• FedDAG is an overloaded acronym representing distinct federated frameworks for domain adversarial generation, clustered learning, and causal DAG structure discovery.
• Each variant employs local training with global coordination to maintain privacy while tackling heterogeneity and unseen domain shifts.
• The frameworks integrate advanced techniques like sharpness-aware aggregation, dual-encoder designs, and acyclic constraint optimization to enhance robust performance.

Searching arXiv for the FedDAG papers and related metadata. FedDAG is an overloaded research acronym used for at least three distinct federated learning frameworks on Federated Domain Adversarial Generation for federated domain generalization in medical image analysis (Che et al., 22 Jan 2025), Federated learning via global DatA and Gradient integration for clustered federated learning under heterogeneous client distributions (Pramanik et al., 26 Feb 2026), and Federated DAG Structure Learning for causal directed acyclic graph discovery from decentralized data under additive noise models (Gao et al., 2021). Although these methods address different problem settings, they share a common federated premise: raw data remain local, while a server coordinates parameter exchange or aggregation to learn a global or partially shared structure.

1. Terminological scope and disambiguation

The term FedDAG does not denote a single canonical method. In the 2025 medical imaging literature, it refers to Federated Domain Adversarial Generation, a framework for federated domain generalization that combines adversarial generation of novel domains with sharpness-aware hierarchical aggregation (Che et al., 22 Jan 2025). In the 2026 clustered FL literature, it refers to Federated learning via global DatA and Gradient integration, a method that integrates data similarity and gradient similarity for client clustering and introduces a dual-encoder architecture for cross-cluster feature transfer (Pramanik et al., 26 Feb 2026). In the 2021 causal discovery literature, it refers to Federated DAG Structure Learning, a gradient-based framework for learning a shared DAG structure from decentralized data while allowing client-specific mechanisms (Gao et al., 2021).

This multiplicity of meanings is substantive rather than merely lexical. Each FedDAG variant is defined by a different objective: out-of-distribution robustness to unseen medical domains (Che et al., 22 Jan 2025), cluster formation and representation sharing in highly heterogeneous FL (Pramanik et al., 26 Feb 2026), or recovery of a shared causal graph under federated constraints (Gao et al., 2021). A plausible implication is that any technical discussion of “FedDAG” requires immediate expansion of the acronym and problem context to avoid conflating incompatible assumptions, architectures, and evaluation protocols.

2. Federated Domain Adversarial Generation

In medical image analysis, FedDAG is a federated domain generalization framework designed to learn a global model from multiple source domains and generalize to an unseen target domain under privacy constraints (Che et al., 22 Jan 2025). The setting assumes $n$ source clients,

$$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$

where each client holds local data that cannot be shared, and the target domain $\mathcal{U}$ is unavailable during training (Che et al., 22 Jan 2025).

The framework is motivated by limitations of prior federated domain generalization methods based on sharing and recombining local domain-specific attributes such as frequency components, feature statistics, or style codes. The reported limitations are threefold: possible privacy concerns, limited coverage because recombination remains inside the global convex hull of source domains, and ambiguous medical domain labels that weaken methods relying on explicit domain annotations (Che et al., 22 Jan 2025). FedDAG instead aims to generate novel-style images different from local and global source domains while preserving semantics (Che et al., 22 Jan 2025).

Its architecture uses a teacher–student–generator triad. On the server, there is a global task model $\mathcal{F}_\theta$ and a global generator $\mathcal{G}_\phi$. On client $i$, the student $\mathcal{S}_{\omega^i}$ is initialized from $\mathcal{F}_\theta$, the teacher $\mathcal{T}_{\theta^i}$ is updated as an EMA of the student, and the generator $\mathcal{G}_{\phi^i}$ is initialized from the global generator (Che et al., 22 Jan 2025). During each communication round, the server distributes task and generator weights; clients perform local training through the NDAG module; clients upload teacher models and generators; and the server performs sharpness-aware evaluation and hierarchical aggregation (Che et al., 22 Jan 2025). A key design choice is that the server never receives student models or data (Che et al., 22 Jan 2025).

The local generation mechanism is perturbation-based rather than generative-from-noise: $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$0 This formulation is intended to preserve content while altering style or appearance (Che et al., 22 Jan 2025). The generator maximizes an instance-level feature discrepancy between the original image’s teacher feature and the generated image’s student feature, while supervised cross-entropy on the generated image preserves semantics. The student is then trained to minimize the same discrepancy and the classification loss, thereby learning domain-invariant representations on generated novel-style samples (Che et al., 22 Jan 2025). The teacher is updated by EMA,

$$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$1

so that the teacher provides a stable anchor for feature matching (Che et al., 22 Jan 2025).

FedDAG further addresses what the paper describes as imbalance in client generalization contributions. Sharpness-aware hierarchical aggregation evaluates local teacher models by perturbing parameters in a SAM-like manner, computing a cross-client validation loss, and using inverse loss as a generalization score $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$2 (Che et al., 22 Jan 2025). These scores are used both for within-client averaging over recent historical models and for across-client weighted aggregation with weights

$$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$3

This design is explicitly intended to bias aggregation toward flatter minima and thereby improve subsequent rounds of adversarial domain generation (Che et al., 22 Jan 2025).

The reported evaluation covers WILDS-Camelyon17, MIDOG2022, GDRBench, and FLamby-ISIC2019, using federated leave-one-domain-out evaluation averaged over five random seeds (Che et al., 22 Jan 2025). Quantitative highlights include AUC improvements over FedAvg from 90.8 to 96.1 on WILDS-Camelyon17, from 76.5 to 80.8 on MIDOG2022, from 70.2 to 78.3 on GDRBench, and from 83.3 to 89.1 on FLamby-ISIC2019 (Che et al., 22 Jan 2025). An ablation on Camelyon17 reports 96.1 AUC for the full method, 93.4 without NDAG, 93.5 without SHA, and 90.8 for FedAvg, indicating that the two components are complementary (Che et al., 22 Jan 2025).

The paper states that this is the first work to introduce such novel domain generation specifically into federated medical scenarios (Che et al., 22 Jan 2025). It also identifies limitations including extra computation from dual-stage local updates, generator communication overhead, sharpness-estimation cost, and untested extension to volumetric 3D imaging (Che et al., 22 Jan 2025).

3. FedDAG for clustered federated learning

In clustered federated learning, FedDAG denotes a framework for heterogeneous non-IID data that combines client clustering with cross-cluster representation sharing (Pramanik et al., 26 Feb 2026). The method addresses four heterogeneity types explicitly listed in the paper: label skew, feature skew, quantity shift, and concept shift (Pramanik et al., 26 Feb 2026). The FL setting partitions $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$4 clients into $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$5 clusters $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$6, with one model $$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$7 per cluster (Pramanik et al., 26 Feb 2026).

The basic cluster model is

$$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$8

while the full FedDAG model uses a dual-encoder architecture

$$\mathcal{D} = \{D_1, D_2, \ldots, D_n\}, \quad D_i = \{(x_k^i, y_k^i)\}_{k=1}^{N^i},$$9

where $\mathcal{U}$0 is a primary encoder specialized to the cluster’s own clients and $\mathcal{U}$1 is a secondary encoder refined using gradients from complementary clusters (Pramanik et al., 26 Feb 2026).

The first major component is a combined similarity metric for clustering. FedDAG computes gradient-based similarity by letting each client perform a warm-up phase without federation, sparsifying its local update $\mathcal{U}$2, and sending the sparse vector $\mathcal{U}$3 to the server. The server then defines pairwise gradient similarity through the cosine angle

$\mathcal{U}$4

Smaller angles indicate more similar local objectives (Pramanik et al., 26 Feb 2026).

The second view is class-wise weighted data similarity. For each client and class, the method performs truncated SVD and sends per-class principal vectors $\mathcal{U}$5 to the server. Pairwise class-wise similarity is measured by principal angles between the class subspaces, with special handling for classes present in only one or neither client (Pramanik et al., 26 Feb 2026). This is then weighted by a class-frequency term,

$\mathcal{U}$6

which is min-max normalized to $\mathcal{U}$7, so that frequency imbalance affects dissimilarity (Pramanik et al., 26 Feb 2026). Client-level data similarity is obtained by averaging these class-wise weighted angles over classes (Pramanik et al., 26 Feb 2026).

FedDAG then fuses the normalized gradient and data similarities with client-specific weights: $\mathcal{U}$8 where $\mathcal{U}$9 is learned by a small MLP through an entropy loss over row-softmaxed adjacency (Pramanik et al., 26 Feb 2026). The paper’s interpretation is that each client can lean more on whichever view yields clearer cluster structure. Clustering itself is performed with agglomerative hierarchical clustering over a threshold search, evaluated by a federated-aware objective combining a compactness term and a degeneracy penalty to discourage tiny clusters (Pramanik et al., 26 Feb 2026). This gives automatic selection of the number of clusters $\mathcal{F}_\theta$0 (Pramanik et al., 26 Feb 2026).

The second major component is Global Representation Sharing via a Cluster Complementarity Graph (CC-Graph) (Pramanik et al., 26 Feb 2026). Complementarity between clusters depends on class demand, class supply, and feature alignment measured through the same class-wise subspace similarities. For each requesting cluster, the method keeps top-$\mathcal{F}_\theta$1 outgoing edges, forming a sparse graph of meaningful cross-cluster transfers (Pramanik et al., 26 Feb 2026).

Training proceeds in two phases. In the primary phase, each cluster trains its primary encoder and classifier using only its own clients’ data while keeping the secondary encoder fixed. In the secondary phase, learner clusters’ secondary encoders are refined by gradients computed on source clusters’ data, as specified by the CC-Graph (Pramanik et al., 26 Feb 2026). The paper emphasizes that primary and secondary updates are independent and can be scheduled flexibly, with the secondary phase performed less frequently to reduce compute (Pramanik et al., 26 Feb 2026).

The reported experiments span CIFAR-10, CIFAR-100, FMNIST, SVHN, CIFAR-10-C, Tiny ImageNet-C, PACS, Office-Caltech-10, and Google Landmarks under several heterogeneity regimes (Pramanik et al., 26 Feb 2026). Under Distribution I with high quantity shift and 20% label skew, the paper reports 90.76% for full FedDAG on CIFAR-10 compared with approximately 42.0% for FedAvg and approximately 85–87% for strong clustered baselines; analogous improvements are reported for FMNIST, SVHN, and CIFAR-100 (Pramanik et al., 26 Feb 2026). Under concept shift, FedDAG reports 69.90% on CIFAR-10, 88.93% on FMNIST, and 85.34% on SVHN, each above FedDAG$\mathcal{F}_\theta$2 and other baselines listed in the paper (Pramanik et al., 26 Feb 2026). For feature skew plus label skew, it reports 65.62% on CIFAR-10-C and 36.27% on Tiny ImageNet-C, again higher than listed baselines (Pramanik et al., 26 Feb 2026). On Google Landmarks, the reported score is 58.23% versus 36.53% for FedAvg and 54.74% for PACFL (Pramanik et al., 26 Feb 2026).

Ablation results distinguish the benefits of clustering from the benefits of cross-cluster sharing. FedDAG$\mathcal{F}_\theta$3 denotes the single-encoder version without the dual-encoder GRS mechanism, while FedDAG$\mathcal{F}_\theta$4 denotes a dual-encoder variant without cross-cluster sharing; the full method outperforms both, which the paper interprets as evidence that the gain comes from structured cross-cluster transfer rather than parameter count alone (Pramanik et al., 26 Feb 2026). The paper also reports that $\mathcal{F}_\theta$5 is sufficient for warm-up, that gradient sparsity of 0.5–1% is enough, and that $\mathcal{F}_\theta$6 is a good trade-off for CC-Graph sparsification (Pramanik et al., 26 Feb 2026).

4. Federated DAG structure learning

The 2021 FedDAG paper addresses causal structure learning from decentralized data under the assumption that each client’s data are generated by an additive noise model (ANM) and that all clients share the same underlying DAG (Gao et al., 2021). For $\mathcal{F}_\theta$7 variables $\mathcal{F}_\theta$8, the ANM is written as

$\mathcal{F}_\theta$9

where the noise variables are independent of their parents and mutually independent across variables (Gao et al., 2021). The paper assumes an Invariant DAG across clients and allows heterogeneity in local mechanisms $\mathcal{G}_\phi$0 or noise distributions (Gao et al., 2021).

The core architectural idea is a two-level local model. Each client has a Graph Structure Learning part, represented by a continuous adjacency proxy $\mathcal{G}_\phi$1, and a Mechanisms Approximating part $\mathcal{G}_\phi$2, a set of neural networks that locally approximate the functional mechanisms among variables (Gao et al., 2021). The structure-learning part is federated and shared through averaging, whereas the mechanism-approximating part remains local in the heterogeneous setting (Gao et al., 2021).

FedDAG uses the continuous acyclicity characterization introduced in NOTEARS,

$\mathcal{G}_\phi$3

but reparameterizes the adjacency matrix through a Gumbel-Sigmoid mask applied to $\mathcal{G}_\phi$4 (Gao et al., 2021). The acyclicity condition becomes

$\mathcal{G}_\phi$5

which is differentiable with respect to $\mathcal{G}_\phi$6 (Gao et al., 2021).

For client $\mathcal{G}_\phi$7, the local score combines a reconstruction term under the masked neural mechanisms and an $\mathcal{G}_\phi$8 sparsity penalty on the mask (Gao et al., 2021). The global optimization problem is to maximize the sum of local scores under the acyclicity constraint with a shared graph parameter $\mathcal{G}_\phi$9 and client-specific mechanism parameters $i$0 (Gao et al., 2021). The equality constraint is handled using an Augmented Lagrangian Method, with outer-loop updates of the multiplier and penalty coefficient and inner-loop federated optimization of the subproblem (Gao et al., 2021).

Within the federated subproblem, each client performs local gradient ascent on its own augmented score, updating both $i$1 and $i$2. Every $i$3 local steps, a subset of clients sends its structural parameters $i$4 to the server, which averages them to obtain $i$5 and broadcasts the result back to all clients (Gao et al., 2021). In the principal heterogeneous setting, only $i$6 is shared; this is the GS-FedDAG variant. In the homogeneous setting, the paper also considers AS-FedDAG, where both structure and mechanism parameters are shared (Gao et al., 2021).

After optimization, the expected Gumbel-Sigmoid mask is thresholded at 0.5 to obtain a hard adjacency, and residual cycles are removed by iteratively deleting the edge with the smallest mask value until acyclicity holds (Gao et al., 2021). The paper also presents identifiability arguments in the decentralized ANM setting, stating that under restricted ANM conditions, minimality, faithfulness, and a shared DAG across clients, the common graph is identifiable from the set of client distributions (Gao et al., 2021).

Experimental evaluation includes synthetic ANM data with Erdős–Rényi and scale-free graphs for $i$7, a variety of function classes including GP, GP-additive, MLP, and MiM, and a real fMRI Hippocampus dataset with 6 brain regions and 7 ground-truth edges (Gao et al., 2021). Metrics include SHD, TPR, NNZ, and FDR (Gao et al., 2021). The reported findings are that GS-FedDAG and AS-FedDAG are close to or better than centralized baselines in homogeneous settings, and that GS-FedDAG is particularly strong in heterogeneous ANM settings where pooled-data baselines degrade due to mechanism mismatch (Gao et al., 2021). On real fMRI data, both GS-FedDAG and AS-FedDAG reportedly outperform PC, NOTEARS, MCSL, GES, and DAG-GNN in SHD relative to anatomical ground truth, and AS-FedDAG achieves SHD $i$8 without raw-data pooling (Gao et al., 2021).

5. Common design patterns and contrasts across the three FedDAG frameworks

Despite their different targets, the three FedDAG frameworks exhibit a recurring decomposition between shared structure and local specialization. In medical FedDAG, the shared components are the global task model and generator, while local adaptation occurs through client-specific student, teacher, and generator updates before sharpness-aware aggregation (Che et al., 22 Jan 2025). In clustered FL FedDAG, each cluster has its own model, but the dual-encoder design separates cluster-specific specialization in the primary encoder from complementary feature import through the secondary encoder (Pramanik et al., 26 Feb 2026). In causal FedDAG, the shared object is the DAG structure parameter $i$9, whereas local mechanism networks $\mathcal{S}_{\omega^i}$0 absorb heterogeneity (Gao et al., 2021).

A second commonality is the use of auxiliary structure beyond naive FedAvg. The medical method augments aggregation with sharpness-based contribution scores and within-client trajectory averaging (Che et al., 22 Jan 2025). The clustered FL method replaces monolithic federation with clustering, adaptive cluster-number selection, and CC-Graph-governed secondary encoder transfer (Pramanik et al., 26 Feb 2026). The causal method imposes an explicit acyclicity constraint and uses augmented Lagrangian optimization instead of unconstrained model averaging (Gao et al., 2021).

The role of heterogeneity also differs. In medical FedDAG, heterogeneity is primarily a domain generalization issue: source hospitals, scanners, staining, and acquisition differences create unseen target-domain shifts (Che et al., 22 Jan 2025). In clustered FL FedDAG, heterogeneity is multi-faceted and explicitly includes label skew, quantity shift, feature skew, and concept shift (Pramanik et al., 26 Feb 2026). In DAG-learning FedDAG, heterogeneity lies in local structural mechanisms or noise distributions under a fixed causal graph (Gao et al., 2021). This suggests that the same acronym names three different strategies for reconciling privacy-preserving collaboration with different forms of non-IID structure.

6. Limitations, misconceptions, and research significance

A common misconception would be to treat FedDAG as a unified line of work. The arXiv record instead shows three independent usages of the same acronym, each embedded in a separate subfield and built around distinct technical assumptions (Che et al., 22 Jan 2025, Pramanik et al., 26 Feb 2026, Gao et al., 2021). Another misconception would be to equate all three with generic federated averaging plus minor modifications. The medical paper adds adversarial perturbation-based domain synthesis and sharpness-aware hierarchical aggregation (Che et al., 22 Jan 2025); the clustered FL paper introduces integrated similarity learning, automatic clustering, and dual-encoder cross-cluster transfer (Pramanik et al., 26 Feb 2026); the DAG-learning paper formulates federated causal discovery as continuous constrained optimization under ANMs (Gao et al., 2021).

Each version also has explicit limitations. The medical method incurs additional generator-related compute, cross-client sharpness evaluation cost, and currently focuses on histopathology, fundus, and dermoscopy benchmarks rather than volumetric imaging (Che et al., 22 Jan 2025). The clustered FL method requires per-class counts and labels, incurs SVD and dual-encoder overhead, and does not implement privacy mechanisms such as differential privacy or secure aggregation, although the paper describes such extensions as straightforward conceptually (Pramanik et al., 26 Feb 2026). The DAG-learning method assumes no latent confounders, relies on ANM assumptions, uses nonconvex acyclicity-constrained optimization with only stationary-point guarantees, and assumes an invariant DAG across clients (Gao et al., 2021).

Taken together, the three FedDAG papers illustrate different trajectories in federated research: robust generalization to unseen medical domains, adaptive grouping and structured sharing under severe non-IIDness, and decentralized causal graph recovery. Their shared importance lies not in a common algorithmic core, but in a common methodological theme: federated systems often benefit from explicitly separating what should be globally coordinated from what should remain client- or cluster-specific.