Federated Domain Adaptation (FedDA)

• FedDA is a family of transfer learning methods that jointly address domain shift and data decentralization while preserving privacy.
• It employs diverse strategies such as adversarial alignment, discrepancy minimization, and domain-aware semantic transfer to optimize adaptation.
• Empirical studies in medical imaging, finance, and vision benchmarks show that FedDA methods enhance accuracy, sensitivity, and communication efficiency.

Federated Domain Adaptation (FedDA) denotes a class of privacy-preserving learning problems in which domain shift and data decentralization are treated jointly: labeled or partially labeled source domains, unlabeled or weakly labeled target domains, or more general domain-skewed clients must collaborate without transmitting raw data. The literature uses the term across several related settings, including multi-source unsupervised domain adaptation under federated constraints, supervised known-domain multi-source adaptation/generalization, cross-modality federated medical segmentation, target-aware aggregation with limited labeled target data, and black-box universal adaptation with inconsistent source label spaces (Peng et al., 2019, Castellon et al., 2023, Le et al., 8 Apr 2026, Sen et al., 15 Sep 2025, Zhou et al., 28 Sep 2025, Liu et al., 2023). This suggests that FedDA is better understood as a family of federated transfer-learning regimes centered on domain shift rather than as a single canonical formulation.

1. Problem formulations and taxonomic scope

A recurrent formulation in FedDA is multi-source unsupervised domain adaptation under federated constraints: multiple source clients hold labeled data, one target client holds unlabeled data, and raw data cannot be centralized. FADA formalizes this as source domains $\mathcal{D}_S=\{\mathcal{D}_{S_i}\}_{i=1}^N$ and one target domain $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$, with the target hypothesis written as a convex combination of source hypotheses, $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$ (Peng et al., 2019). FedDaDiL uses essentially the same multi-source structure, but treats each client distribution as an empirical distribution and the target client as unlabeled (Castellon et al., 2023).

A second line of work broadens the setting beyond classical unsupervised target adaptation. FedDAP addresses a supervised multi-domain federated setting with known domain identities and shared label space, where clients “share a same label space $\mathcal{Y}$ (i.e., $P(y)$ is consistent) while their feature distributions $P(x)$ differ,” and domain adaptation occurs through domain-aware prototypes rather than a separate unlabeled target silo (Le et al., 8 Apr 2026). FedDAF studies one target client “having only a small amount of target data,” and adapts a source global model to that target through target-aware aggregation, using

$w = \beta w^S + (1-\beta)w^T$

as the generic FDA fusion form (Sen et al., 15 Sep 2025).

Other formulations are more specialized. The medical-segmentation work titled “Federated Domain Adaptation” studies a setting where clients may hold disjoint image modalities, formalized by

$\exists i,j \in (0,N),m_i\cap m_j=0,$

and seeks a single segmentation model that performs across the union of all modalities (Zhou et al., 28 Sep 2025). PrADA addresses a cross-silo regime in which the target-domain party is insufficient in both samples and features, combining vertical federated learning with adversarial domain adaptation (Kang et al., 2021). UFDA further relaxes assumptions by allowing inconsistent source label spaces and a blind target label space, while exposing only black-box source models and source label-set information (Liu et al., 2023). ST-LF departs from fixed roles entirely by jointly deciding which decentralized devices should act as sources and which as targets in a multi-source to multi-target setting (Wang et al., 2023).

Regime	Defining characteristic	Representative papers
Multi-source federated UDA	labeled source clients, unlabeled target client	(Peng et al., 2019, Castellon et al., 2023)
Supervised known-domain FedDA/FedDG	domain-tagged clients, shared label space	(Le et al., 8 Apr 2026)
Target-aware low-label FDA	one target client with limited labeled target data	(Sen et al., 15 Sep 2025)
Cross-modality federated segmentation	clients hold different image modalities/domains	(Zhou et al., 28 Sep 2025)
Vertical cross-silo FedDA	target insufficient in both samples and features	(Kang et al., 2021)
Universal black-box FedDA	inconsistent source label spaces, blind target label space	(Liu et al., 2023)
Decentralized multi-source to multi-target	source-target determination and link formation	(Wang et al., 2023)

This diversity of formulations is not a mere terminological variation. It changes what counts as a source, what information is exchanged, whether the target participates during training, whether label spaces are shared, and whether adaptation is framed as feature alignment, prototype transfer, hypothesis aggregation, or decentralized role assignment.

2. Core optimization paradigms

One major FedDA paradigm is adversarial alignment. FADA extends adversarial adaptation to the federated setting through source-specific feature generators $G_i$, a target generator $G_t$, and pairwise domain identifiers $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$0, with source-target alignment written as

$\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$1

and target-source importance controlled by dynamic attention weights $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$2 (Peng et al., 2019). FedMM studies the same adversarial DA family from the optimizer side. It reformulates federated adversarial DA as a consensus-constrained minimax problem,

$\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$3

and introduces an augmented-Lagrangian primal-dual method to control client drift under label imbalance and target-only clients (Shen et al., 2021). In cross-modality segmentation, adversarial alignment is applied directly to feature maps, with FedDA-cyclic and FedDA-joint differing only in how target feature maps are obtained (Zhou et al., 28 Sep 2025).

A second paradigm is discrepancy minimization in compact statistics or transport space. UG-FedDA aligns domain-level style distributions by modeling source and target style priors as Gaussians and minimizing a closed-form KL divergence,

$\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$4

where uncertainty estimated by MC dropout down-weights unreliable features (Zhu et al., 5 Dec 2025). FedRF-TCA starts from TCA and RF-TCA, then redesigns the alignment objective for FDA so that clients exchange only compressed random-feature statistics $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$5, giving communication complexity independent of sample size according to the abstract (Feng et al., 2023). The FedDaDiL line replaces invariant-feature matching by Wasserstein barycentric dataset dictionary learning: each client distribution $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$6 is represented as a barycenter $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$7 of shared atoms, while client-specific barycentric coordinates remain private (Castellon et al., 2023, Montesuma et al., 2024). De-FedDaDiL eliminates the server and reaches consensus through peer-to-peer atom exchange and local aggregation (Clain et al., 22 Mar 2025).

A third paradigm is domain-aware semantic transfer. FedDAP constructs one global prototype for each class-domain pair $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$8 rather than a single global prototype per class, with local training combining domain-consistent prototype alignment and cross-domain prototype contrastive learning (Le et al., 8 Apr 2026). MPFT likewise rejects averaging as the central operation, but does so through multi-domain prototypes extracted from a frozen pretrained encoder and a server-side supervised adapter trained on the union of those prototypes (Zhang et al., 2024). FedDAF belongs to a related but distinct line: it does not align features directly, but computes a target-aware aggregation weight from the angle between source and target mean gradient fields, normalized with a Gompertz function (Sen et al., 15 Sep 2025).

These paradigms encode different views of what must be transferred. Adversarial methods seek domain-invariant or modality-invariant internal representations. Transport and discrepancy methods manipulate domain summaries, atoms, or kernelized statistics. Prototype and target-aware aggregation methods treat the transferable object as structured semantic information or target-objective-aligned model behavior.

3. Architectural patterns and communicated objects

FedDA methods differ sharply in what they expose to the federation. UG-FedDA is architecturally homogeneous across clients and communicates model parameters together with domain-level style statistics $\mathcal{D}_T=\{x_j^t\}_{j=1}^{n_t}$9, not raw features or raw samples (Zhu et al., 5 Dec 2025). FedDAP communicates local class-domain prototypes and domain IDs, while keeping raw data and per-sample features local (Le et al., 8 Apr 2026). FedDaDiL and De-FedDaDiL transmit updated atom dictionaries but keep client-specific barycentric coordinates $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$0 private, because those coordinates are precisely the variables that identify how to reconstruct a given client’s data distribution from the dictionary (Castellon et al., 2023, Clain et al., 22 Mar 2025). MPFT reduces communication to one prototype upload plus adapter redistribution, replacing iterative model exchange by one-round prototype-mediated fine-tuning (Zhang et al., 2024). PrADA exchanges encrypted high-order features and cut-layer gradients under a PHE-based protocol rather than ordinary model updates (Kang et al., 2021).

Architectural design is equally heterogeneous. UG-FedDA builds its style extractor on 1D convolution, Transformer blocks, patch merging, and adaptive average pooling over multi-template MRI ROI tokens, then combines domain-level style features with sample features in an evidential classifier (Zhu et al., 5 Dec 2025). FDAC uses a ViT backbone and explicitly manipulates hidden block outputs $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$1 for model-contrastive domain augmentation, treating source-model-induced latent views of target samples as positives in a contrastive objective (Yi et al., 2023). MPFT freezes a pretrained CLIP image encoder and text-derived classification head, and trains only a global adapter on uploaded prototypes (Zhang et al., 2024). PrADA assigns one extractor, one discriminator, and one aggregator to each feature group, reflecting its emphasis on semantically meaningful high-order features and group-wise adaptation (Kang et al., 2021).

The following table summarizes the principal communication objects.

Framework	Exchanged object	Kept local
UG-FedDA	model parameters and domain-level style statistics/uncertainty summaries	raw scans, raw features (Zhu et al., 5 Dec 2025)
FedDAP	local prototypes and domain IDs	raw data (Le et al., 8 Apr 2026)
FedDaDiL / De-FedDaDiL	atom dictionaries	raw data, barycentric coordinates $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$2 (Castellon et al., 2023, Clain et al., 22 Mar 2025)
MPFT	prototypes and global adapter	raw data (Zhang et al., 2024)
PrADA	encrypted high-order features and secure protocol messages	raw features, true $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$3 (Kang et al., 2021)
FedDA-cyclic	feature maps plus segmentation parameters	raw images (Zhou et al., 28 Sep 2025)

This variation has methodological consequences. Methods that exchange domain summaries or prototypes can reduce communication cost and expose semantically structured objects. Methods that exchange feature maps or encrypted activations can support fine-grained alignment but raise sharper questions about privacy leakage and protocol complexity.

4. Representative systems and application domains

Medical imaging has become a prominent FedDA application domain. UG-FedDA studies multicenter Alzheimer’s disease classification from baseline T1-weighted structural MRI across ADNI, AIBL, and OASIS, alternating each center as target and reporting strong cross-domain performance. For NC vs AD, the abstract reports accuracies of 90.54%, 89.04%, and 77.78% on ADNI, AIBL, and OASIS, respectively; for MCI vs AD, 80.20%, 71.91%, and 79.73%; and for NC vs MCI, 76.87%, 73.91%, and 83.73% (Zhu et al., 5 Dec 2025). The federated segmentation framework titled “Federated Domain Adaptation” addresses a more severe setting in which some hospitals retain only MRI scans while others retain only CT scans. In the two-client cardiac MMWHS setting, mean Dice improves from 80.2 with FedAvg to 84.3 with FedDA-cyclic and 84.8 with FedDA-joint; in the two-client abdominal CHAOS setting, mean Dice rises from 57.1 with FedAvg to 65.1 with FedDA-cyclic (Zhou et al., 28 Sep 2025). PrADA, although centered on tabular finance rather than images, shows that FedDA is not restricted to vision: on Loan Default with 40 positive labels, full PrADA outperforms SecureBoost by 17.87% AUC and 28.83% KS (Kang et al., 2021).

Visual classification benchmarks provide the broadest comparative picture. FADA reports average accuracies of 73.6 on Digit-Five, 87.1 on Office-Caltech10 with ResNet101, 30.3 on DomainNet with ResNet101, and 78.9 on Amazon Review for its full model (Peng et al., 2019). FedDaDiL reaches 84.11 and 84.17 average accuracy on TEP, 97.96 and 97.95 on CWRU, and 94.42 and 94.16 on Caltech-Office 10 for its R and E variants, respectively (Castellon et al., 2023). FDAC reports average accuracies of 80.8 on OfficeHome, 98.5 on OfficeCaltech, 93.4 on PACS, 53.8 on DomainNet, and 95.5 on a breast-cancer dataset (Yi et al., 2023). FedDAP, in a supervised multi-domain setting, attains 65.20 on DomainNet, 72.53 on Office-10, and 84.63 on PACS, while also reporting leave-one-domain-out target accuracies of 49.52 on DomainNet and 65.70 on Office-10 (Le et al., 8 Apr 2026). MPFT, which emphasizes one-round communication, reports on DomainNet subset-50 ood acc / ind acc of 0.8077 / 0.7813 for average sampling and 0.8236 / 0.8294 for random sampling at rate 0.3 (Zhang et al., 2024).

Several works are especially valuable because they widen the operational envelope of FedDA rather than only improving accuracy. ST-LF addresses multi-source to multi-target decentralized transfer and reports 71.32% accuracy with 21.10% normalized energy on MNIST, 60.23% with 22.20% energy on USPS, and 36.44% with 27.70% energy on MNIST-M (Wang et al., 2023). UFDA demonstrates that adaptation remains possible even when the target sees only black-box source APIs and source label-set information, with average accuracies of 77.18 on Office-Home, 89.28 on Office-31, and 75.41 on the VisDA+ImageCLEF-DA setup for HCLD$h_T := \sum_{i=1}^N \alpha_i h_{S_i}$4 (Liu et al., 2023). FedRF-TCA argues, at the abstract level, that communication complexity can be made independent of sample size while remaining robust to network condition (Feng et al., 2023).

5. Empirical regularities and comparative lessons

A consistent empirical lesson is that explicit adaptation improves cross-domain generalization beyond plain federated aggregation. UG-FedDA directly isolates this effect: on ADNI, adding domain adaptation changes NC vs AD ACC from 91.22% to 90.54%, but improves sensitivity from 77.55% to 83.67% and AUC from 91.03% to 91.34%; for MCI vs AD, ACC rises from 76.24% to 80.20% and sensitivity from 61.11% to 83.33% (Zhu et al., 5 Dec 2025). In the segmentation setting, unilateral adversarial learning improves over FedAvg, but the best results require both clients to participate, indicating that mutual adversarial interaction is essential under modality discrepancy (Zhou et al., 28 Sep 2025). FADA’s ablations show the same structure: on Digit-Five, average accuracy moves from 61.3 for FADA + attention to 70.3 after adding adversarial alignment and 73.6 after adding disentanglement (Peng et al., 2019).

A second regularity is that domain-aware structure outperforms domain-agnostic averaging. FedDAP’s ablations show that neither domain-consistent prototype alignment nor cross-domain prototype contrastive learning alone is sufficient; on DomainNet, Office-10, and PACS, using both yields 65.20 / 72.53 / 84.63, compared with 62.61 / 68.53 / 78.74 for DPA alone and 62.86 / 62.18 / 81.87 for CPCL alone (Le et al., 8 Apr 2026). MPFT reaches the same conclusion from a different angle: richer multi-domain prototype sampling beats both average-sampling MPFT and iterative FL baselines, while converging in 1 communication round (Zhang et al., 2024). The FedDaDiL papers make a related point by showing that modeling domains as Wasserstein barycenters can outperform adversarial and pseudo-label baselines, and remain robust when the number of local client iterations $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$5 grows (Montesuma et al., 2024, Castellon et al., 2023).

A third lesson concerns uncertainty, target awareness, and communication robustness. FedDAF reports its largest gains when target labels are especially scarce: on CIFAR-10 with target fraction 0.05 and noise 0.3, FedDAF reaches 58.25 versus 41.35 for FedGP (Sen et al., 15 Sep 2025). UG-FedDA shows that uncertainty is not merely a confidence report: uncertainty-weighted aggregation improves NC vs AD on ADNI from 89.19% ACC and 69.39% SEN under FedAvg to 90.54% and 83.67% (Zhu et al., 5 Dec 2025). FedMM shows that optimizer design matters under adversarial FedDA: when training from scratch it outperforms other GDA-based federated average methods by around 20% in accuracy over the same communication rounds, and when training from pre-trained models it improves by 5.4% to 9% for different networks (Shen et al., 2021). FedRF-TCA generalizes the communication argument by claiming sample-size-independent communication complexity and robustness to network condition (Feng et al., 2023).

These empirical patterns suggest that the most successful FedDA systems do not treat federation as a thin privacy wrapper around centralized DA. They typically redesign the transferable object—style statistics, atoms, prototypes, mean gradient fields, random-feature summaries, or adapter training sets—so that adaptation and communication constraints are coupled from the outset.

6. Limitations, controversies, and open directions

A first limitation is setting fragmentation. The term FedDA covers classical multi-source unsupervised domain adaptation, supervised known-domain representation learning, target-aware low-label adaptation, universal black-box adaptation, cross-modality segmentation, and even vertical cross-silo transfer (Peng et al., 2019, Le et al., 8 Apr 2026, Sen et al., 15 Sep 2025, Liu et al., 2023, Kang et al., 2021). This suggests that comparisons across papers can be misleading when source/target roles, label availability, and domain observability differ. Several papers state this explicitly: FedDAP is “not classical unsupervised FedDA,” while FedFD-A is better characterized as federated domain generalization with test-time adaptation rather than true training-time FedDA (Le et al., 8 Apr 2026, Yang et al., 2024).

A second limitation is assumption load. Many methods assume a shared label space, known domain identities, a single target client, or target data availability during training. FedDAP assumes known domain identifiers and supervised labels on all clients (Le et al., 8 Apr 2026). FedDAF studies only one target client and does not analyze unlabeled-only target use (Sen et al., 15 Sep 2025). UG-FedDA fixes a two-source/one-target topology over three known MRI cohorts (Zhu et al., 5 Dec 2025). ST-LF requires divergence and discrepancy estimates as well as a nontrivial decentralized optimization layer (Wang et al., 2023). PrADA assumes privacy-preserving entity alignment and expert-defined feature groups (Kang et al., 2021).

A third limitation concerns privacy claims. Many works are privacy-preserving only in the standard FL sense that raw data remain local. FADA does not provide cryptographic guarantees and notes that sharing features may still leak information (Peng et al., 2019). FedDaDiL and De-FedDaDiL keep barycentric coordinates private but do not provide DP or secure aggregation (Montesuma et al., 2024, Clain et al., 22 Mar 2025). The segmentation FedDA paper argues that feature-map transfer “has little impact on privacy,” but states this as an assumption rather than a formal guarantee (Zhou et al., 28 Sep 2025). MPFT applies additive Gaussian noise to prototypes,

$h_T := \sum_{i=1}^N \alpha_i h_{S_i}$6

and reports that raw image reconstruction remains unsuccessful even after one million gradient descent iterations, but it does not provide an $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$7-DP accountant (Zhang et al., 2024). PrADA offers the strongest protocol-level treatment through PHE and masking, yet explicitly notes that its PP-VFL does not inherently prevent label inference attacks because the cut-layer gradient $h_T := \sum_{i=1}^N \alpha_i h_{S_i}$8 is revealed to party C after decryption (Kang et al., 2021).

Finally, scalability and generality remain open. The FedDA literature repeatedly identifies unresolved issues: scalability to many clients or continually arriving sites, mixed or latent domains within a client, graph-aware decentralized communication, stronger calibration analysis for uncertainty-based methods, and richer theoretical guarantees for nonconvex adversarial or prototype-based training (Zhu et al., 5 Dec 2025, Clain et al., 22 Mar 2025, Sen et al., 15 Sep 2025, Shen et al., 2021). A plausible implication is that future progress will depend less on introducing yet another alignment penalty than on clarifying which FedDA regime is being solved, which information is allowed to move, and how privacy, communication, and target risk should be traded off in that regime.