Federated Source-Free Domain Adaptation

• Federated Source-Free Domain Adaptation (FFREEDA) is a framework that adapts a server-trained source model to multiple clients with only unlabeled, heterogeneous target data.
• It combines federated learning with source-free domain adaptation to address challenges like pseudo-label noise, client drift, and inter-client domain shifts.
• Methodological patterns such as clustering, alignment regularization, and representation-first designs enable privacy-preserving adaptation in tasks like semantic segmentation and classification.

Searching arXiv for papers on Federated Source-Free Domain Adaptation to ground the article in the cited literature. Federated Source-Free Domain Adaptation, written in the literature as FFREEDA or FFreeDA, denotes a federated learning setting in which a server first trains a model on labeled source-domain data, then adapts that model collaboratively across multiple clients that hold only unlabeled target-domain data, while the source dataset is unavailable during adaptation. The setting combines federated learning, source-free domain adaptation, and unsupervised client-side training, and it has been studied in both semantic segmentation and classification. Its central difficulty is that the system must transfer a source-pretrained model across heterogeneous target clients without revisiting source samples and without access to target labels, so pseudo-label quality, client drift, and cross-client domain heterogeneity become first-order concerns (Shenaj et al., 2022, Mori et al., 2024, Yashwanth et al., 7 Dec 2025).

1. Problem formulation and scope

A standard FFREEDA formulation begins with a labeled source dataset on the server and unlabeled target datasets on clients. One classification-oriented definition writes the source dataset as

$D_s = \{(x_s^i, y_s^i)\ |\ i=1,\ldots,N_s\},$

and the target data on each of $K$ clients as

$$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$

The workflow is: the server trains a source model using labeled source data, broadcasts that pretrained source model to clients, each client adapts the model using its unlabeled target-domain data, and the server aggregates client updates to obtain one global model adapted to the target domain. In the same literature, the source model is often decomposed as $f^s = h^s \circ g^s$, with a feature extractor $g^s$ and classifier $h^s$, and the client distributions are explicitly heterogeneous, with

$\mathbb{P}_{k'}(x) \neq \mathbb{P}_{l'}(x)$

for different clients or domains (Kihara et al., 10 Sep 2025, Yashwanth et al., 7 Dec 2025).

In semantic segmentation, the task is framed analogously but with pixel-wise outputs and client-side unlabeled image collections. One formal objective is

$$w^* = \arg\min_w \sum_{k\in[K]} \frac{|\mathcal{D}_k^T|}{|\mathcal{D}^T|}L_k(w),$$

where $\mathcal{D}^T=\bigcup_{k \in K} \mathcal{D}^T_k$. In that formulation, the source and target share the same semantic label set, but target clients are non-iid and may be drawn from latent visual domains (Shenaj et al., 2022).

The scope of the setting has expanded. Early work introduced FFREEDA for semantic segmentation under server-only source pre-training and unlabeled clients. Later work extended FFREEDA to classification, emphasizing that prior FFREEDA work had focused only on semantic segmentation. Subsequent studies further specialized the setting to Class Imbalanced FFREEDA (CI-FFREEDA), where class imbalance appears in the source domain, in target domains, and across target clients, and to multi-viewpoint semantic segmentation with car and drone clients under adverse weather (Shenaj et al., 2022, Mori et al., 2024, Kihara et al., 10 Sep 2025, Rizzoli et al., 2024).

2. Core technical difficulties

FFREEDA is difficult because no labeled source data are available during adaptation, clients are unlabeled, and large domain gaps can exist both between source and target and across clients. In source-free adaptation, clients therefore rely on pseudo-labels; under large domain shifts, these pseudo-labels become unreliable, and in federated training poor pseudo-labels cause client drift, where local models overfit their own biased pseudo-label distributions and aggregated updates become suboptimal. This challenge is explicit in classification work that identifies the core problem as adapting a pretrained server model to multiple heterogeneous unlabeled client domains, without source data, while preventing pseudo-label collapse and client drift (Yashwanth et al., 7 Dec 2025).

A second difficulty is that heterogeneity is not limited to covariate shift. CI-FFREEDA introduces intra-domain class imbalance, inter-domain label shift, and inter-client label shift, with relations such as

$p(y_s) \neq q(y_t),$

and, in the federated case,

$K$0

That extension is presented as more realistic than standard FFREEDA because it combines source imbalance, target imbalance, source-target label shift, client-to-client label shift, and domain gap (Kihara et al., 10 Sep 2025).

A third difficulty arises in semantic segmentation when heterogeneity reflects not only domain appearance but also sensing geometry and environmental conditions. The adverse-weather, multi-viewpoint formulation includes both car and drone clients, with client populations that are unlabeled, non-i.i.d., unbalanced, and semantically or viewpoint heterogeneous. The target side is partitioned across clients with disjoint local datasets, while weather varies dynamically during training. This setting compounds standard source-free self-training instability with viewpoint shift and weather shift (Rizzoli et al., 2024).

These formulations collectively establish that FFREEDA is not merely federated learning with missing labels, nor merely source-free domain adaptation distributed across devices. It is the joint problem of privacy-preserving adaptation, unlabeled multi-client optimization, and severe statistical heterogeneity (Shenaj et al., 2022).

3. Dominant methodological patterns

A common baseline pattern is local adaptation with pseudo-labeling followed by federated aggregation. In one classification formulation, target pseudo-labels are generated from class prototypes,

$K$1

and the local loss combines Information Maximization and pseudo-label cross-entropy. In its federated version, each client independently optimizes that loss and the server aggregates via FedAvg. This baseline is termed FedLoA, and it is described as weak in FFREEDA because each client self-trains on noisy pseudo-labels, pseudo-label noise differs across domains, local optimization drifts toward client-specific minima, and aggregating drifting client models harms the global model (Yashwanth et al., 7 Dec 2025).

A second methodological pattern is clustering and personalization. In semantic segmentation, LADD combines source pre-training enhanced by style transfer, local self-training with regularization, and a federated clustered aggregation scheme based on clients’ style. The model is split into global parameters $K$2 and cluster-specific parameters $K$3, with global aggregation for $K$4 and cluster-specific aggregation for $K$5. In classification, FedWCA replaces style exchange with private, parameter-free clustering using model parameters, specifically the first-layer parameters of each client’s feature extractor, and then performs weighted aggregation across cluster models rather than strict within-cluster averaging (Shenaj et al., 2022, Mori et al., 2024).

A third pattern is consistency and alignment regularization. FedSCAl keeps a local adaptation backbone but adds Server-Client Alignment (SCAl), a regularizer that aligns client and server predictions. SCAl has two terms: a client alignment loss that matches weakly and strongly augmented client predictions under a confidence threshold, and a server alignment loss that matches client predictions to the server model’s confident predictions. The full local client objective is

$K$6

This design uses the server’s pretrained model as a stabilizing teacher signal during federated adaptation, without source data (Yashwanth et al., 7 Dec 2025).

A fourth pattern is representation-first design. The CI-FFREEDA study argues that many existing methods are improving the wrong part of the pipeline and proposes replacing the conventional CNN backbone with a frozen vision foundation model, specifically DINOv2, while training and aggregating only the bottleneck and classifier. The paper’s central claim is that the backbone matters more than the choice of FFREEDA algorithm, and it uses existing SFDA methods such as SHOT, NRC, AaD, ISFDA, and ICPR largely as heads on top of a stronger frozen representation (Kihara et al., 10 Sep 2025).

Paper	Task	Main mechanism
LADD (Shenaj et al., 2022)	Semantic segmentation	Style-driven clustered aggregation, KD, SWA teacher
FedWCA (Mori et al., 2024)	Classification	Parameter-free clustering, weighted cluster aggregation, improved pseudo-labeling
HyperFLAW (Rizzoli et al., 2024)	Semantic segmentation	Weather-aware BN, hyperbolic prototypes, server smoothing
FedSCAl (Yashwanth et al., 7 Dec 2025)	Classification	Server-client alignment with adaptive thresholding
CI-FFREEDA VFM study (Kihara et al., 10 Sep 2025)	Classification under class imbalance	Frozen DINOv2 backbone, balanced sampling

4. Representative algorithms

LADD introduced FFREEDA as a more realistic semantic-segmentation setting in which dense pixel annotations are unavailable on clients and the server can use labeled source data only during pre-training. Its architecture uses FDA-style Fourier amplitude statistics as compact style representations, clusters clients according to style vectors by selecting the best clustering with silhouette score, and maintains one global parameter set plus cluster-specific layers. Local training combines pseudo-label self-training, knowledge distillation from the pre-trained model, and an SWA-style teacher averaging mechanism: $K$7 At test time, the style of a test image is extracted, compared to cluster centroids by L2 distance, and the corresponding cluster model is used (Shenaj et al., 2022).

FedWCA extends FFREEDA to classification and defines the goal as learning personalized models $K$8 for each client. Its first phase clusters clients using only model parameters and the FINCH procedure. Its second phase builds a weighted mixture of cluster models through client-specific weights $K$9 and pseudo-performance weights $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$0, producing an initial model

$$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$1

Its third phase performs source-free local adaptation using a SHOT-style objective, prototype-based pseudo-labeling, fixed pseudo-labels within each communication round, two-model pseudo-label selection, and a mixup step for mismatched pseudo-labels with $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$2 in experiments (Mori et al., 2024).

FedSCAl addresses the observation that simple federated adaptation of source-free methods struggles with client drift in real-world scenarios due to extreme data heterogeneity. At each communication round, the server sends current global weights to selected clients, clients train locally for $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$3 epochs, return updated parameters, and the server aggregates them. The distinguishing component is SCAl. For client alignment, weak-augmented predictions are converted to one-hot pseudo-labels and regularize strong-augmented predictions via KL divergence under a confidence threshold $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$4. For server alignment, the one-hot prediction of the server model on weak augmentation supervises the client’s strong-augmentation prediction, again gated by confidence. The method also proposes adaptive thresholding based on prediction entropy, using

$$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$5

and

$$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$6

The stated interpretation is that positive skew indicates more confident predictions and tightens the threshold, whereas negative skew relaxes it (Yashwanth et al., 7 Dec 2025).

HyperFLAW specializes FFREEDA to semantic segmentation with car and drone clients under clear day, night, rain, and fog. The method combines pseudo-label self-training with prototype clustering,

$$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$7

a weather-aware batch-normalization strategy that keeps separate BN parameters for each weather condition, and prototype alignment in hyperbolic space using the Poincaré ball model. Curvature is treated as learnable and updated on clients with Riemannian Adam, while global model aggregation is stabilized by averaging the current round’s federated model with a queue of previous global models (Rizzoli et al., 2024).

The CI-FFREEDA VFM study departs from loss-centric and aggregation-centric design. It keeps the standard model form $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$8, replaces $$D_t^{(k)} = \{x_t^{(k),i}\ |\ i=1,\ldots,N_t^{(k)}\}.$$9 with a frozen DINOv2 backbone, freezes backbone parameters during both source training and target adaptation, trains only the bottleneck and classifier, and uses balanced sampling in source training. It explicitly notes that, unlike some prior SFDA methods that freeze the classifier, the classifier remains trainable because a fixed backbone can reduce bottleneck expressiveness, especially when imbalance ratios vary across clients (Kihara et al., 10 Sep 2025).

5. Experimental regimes and reported performance

The empirical literature spans both segmentation and classification, with differing metrics and benchmark conventions. Semantic-segmentation papers report mIoU, while classification papers report classification accuracy (%) or macro averaged recall (MAR). Evaluation settings include GTA5 as source with Cityscapes, CrossCity, and Mapillary as targets for segmentation; Office-Home, DomainNet, Office-31, Digit-Five, PACS, and VisDA for classification; and additional adverse-weather evaluation on ACDC and FLYAWARE-R (Shenaj et al., 2022, Rizzoli et al., 2024, Mori et al., 2024, Kihara et al., 10 Sep 2025, Yashwanth et al., 7 Dec 2025).

For semantic segmentation, LADD reported the following representative results. On GTA5 $f^s = h^s \circ g^s$0 Cityscapes, Source only achieved $f^s = h^s \circ g^s$1, FedAvg + Self-Tr. $f^s = h^s \circ g^s$2, and LADD $f^s = h^s \circ g^s$3 or $f^s = h^s \circ g^s$4 depending on the variant. On GTA5 $f^s = h^s \circ g^s$5 CrossCity, Source only achieved $f^s = h^s \circ g^s$6, MCD federated $f^s = h^s \circ g^s$7, FedAvg + Self-Tr. $f^s = h^s \circ g^s$8, and LADD $f^s = h^s \circ g^s$9 or $g^s$0. On GTA5 $g^s$1 Mapillary, Source only achieved $g^s$2, FedAvg + Self-Tr. $g^s$3, and LADD $g^s$4 or $g^s$5. The strongest gains were reported on the more heterogeneous CrossCity and Mapillary setups (Shenaj et al., 2022).

HyperFLAW evaluated a harder benchmark with car and drone clients in adverse weather. On the main benchmark, Source Only achieved 22.70 mIoU, FedAvg+ST 26.10, LADD 26.96, and HyperFLAW 28.96. On the larger ACDC+FA-R-XL setup, HyperFLAW again outperformed the baselines, reaching 28.84 mIoU versus 26.86 for LADD and 24.59 for FedAvg+ST. The same paper introduced FLYAWARE-S with 108,180 generated images and FLYAWARE-R with 1,287 domain-translated images (Rizzoli et al., 2024).

For classification, FedWCA reported mean accuracy across clients and target domains over 15 trials. Its average accuracies were 77.86\% on Digit-Five, 78.38\% on PACS, and 65.97\% on Office-Home, outperforming Source Only, Local, FedAvg, LADD, FedPCL+PL, and extended FL baselines such as FedProx, FedAMP, and IFCA (Mori et al., 2024).

FedSCAl evaluated a cross-silo setup with an ImageNet-pretrained ViT-Small backbone on Office-Home, DomainNet, and Office-31. On Office-Home, examples include Source = Art with FedLoA average 75.55 versus FedSCAl 79.05, Source = Clipart with 83.18 versus 86.01, Source = Product with 71.94 versus 75.56, and Source = Real-World with 73.08 versus 76.59. The reported gains over FedLoA were +3.50\% for Art, +3.51\% for Real-World, +2.83\% for Clipart, and +3.62\% for Product. On DomainNet, examples include Quickdraw-pretrained with FedLoA 48.24 and FedSCAl 60.61, and Real-pretrained with 70.15 and 78.08. On Office-31, Amazon-pretrained yielded 96.15 for FedLoA and 96.46 for FedSCAl, DSLR-pretrained 88.39 and 89.25, and Webcam-pretrained 88.56 and 90.76. An ablation on Office-Home reported averages of 75.93 with neither alignment term, 77.46 with client alignment only, 77.66 with server alignment only, and 79.30 with both (Yashwanth et al., 7 Dec 2025).

The CI-FFREEDA VFM study reported that replacing a ResNet backbone with frozen DINOv2 features substantially changed outcomes. On Office-Home, averaged over source-target pairs, ResNet-50 achieved 63.2 MAR, DINOv2 ViT-S 73.7 MAR, and DINOv2 ViT-B 79.8 MAR, corresponding to gains of +10.5 and +16.6. On VisDA, averaged over methods, ResNet-101 achieved 73.2, while ViT-S / ViT-B achieved 77.4, a gain of +4.2. The same study also quantified robustness to domain gaps and imbalance: in domain adaptation without class imbalance, the source-to-target drop was -17.6 for ResNet-50, -10.6 for ViT-S, and -7.3 for ViT-B; under inter-domain label shift, the drop from source performance to imbalanced target performance was -21.8, -16.0, and -11.8, respectively (Kihara et al., 10 Sep 2025).

6. Limitations, contested interpretations, and practical implications

A persistent disagreement in the FFREEDA literature concerns the primary bottleneck. FedSCAl, FedWCA, LADD, and HyperFLAW each improve performance through better aggregation, clustering, alignment, or stabilization mechanisms. By contrast, the CI-FFREEDA work argues that existing SFDA methods and FL aggregation tricks help less than expected under real-world imbalance, and that the feature extractor is the main lever. This does not invalidate algorithmic advances; rather, it establishes a live research tension between representation strength and adaptation or federation design (Kihara et al., 10 Sep 2025, Yashwanth et al., 7 Dec 2025, Mori et al., 2024).

Several limitations are explicit. FedSCAl assumes a predefined number of classes across clients, which may reduce applicability when class sets evolve or are unknown, and it still depends on a reasonably useful source-pretrained server model, confidence thresholding, and pseudo-label quality. HyperFLAW remains challenged by severe conditions such as night, relies on a weather classifier plus curated weather-specific BN branches, and points toward multimodal data and more advanced weather adaptation strategies as future work. LADD’s clustering depends on exchanged style representations, whereas FedWCA presents its parameter-based clustering partly as a response to the privacy concerns raised by style feature sharing (Yashwanth et al., 7 Dec 2025, Rizzoli et al., 2024, Shenaj et al., 2022, Mori et al., 2024).

Practical considerations are equally important. FedSCAl reports only a modest extra computation cost due to additional alignment forward passes and unchanged communication cost because only model parameters are exchanged. The frozen-VFM approach pushes this further: because the backbone is frozen, only the bottleneck and classifier are updated and transmitted, and those together are less than 1 MB, compared with roughly 94 MB for ResNet-50 and 169 MB for ResNet-101. The same study reports 11.0 G FLOPs for frozen ViT-S, 43.9 G FLOPs for frozen ViT-B, and, with feature banking, 0.61 M FLOPs for ViT-S and 1.20 M FLOPs for ViT-B. This suggests that some FFREEDA deployments may become more practical not by increasing adaptation complexity but by freezing a stronger backbone (Yashwanth et al., 7 Dec 2025, Kihara et al., 10 Sep 2025).

Across these variants, a common conclusion emerges: FFREEDA is a realistic benchmark for privacy-preserving adaptation because the source domain is unavailable during target training, client data are unlabeled, and target distributions are heterogeneous. The literature diverges on whether the most effective remedy is clustered personalization, alignment-based regularization, weather- and geometry-aware adaptation, or stronger pretrained representations. A plausible implication is that future FFREEDA systems will increasingly combine these strands rather than treat them as mutually exclusive.