---
title: Active Source-Free Domain Adaptation
url: https://www.emergentmind.com/topics/active-source-free-domain-adaptation-asfda
type: topic
---

# Active Source-Free Domain Adaptation

Active Source-Free Domain Adaptation (ASFDA) is a domain adaptation setting in which a source-trained model is transferred to a target domain without access to source data, while a small target-domain annotation budget is used through active selection of the most valuable target samples. In the original formulation, a source-trained model \(h_s(x)=g_s(f_s(x))\) is given together with an unlabeled target pool \(\mathcal D_t=\{x_i\}_{i=1}^{N_t}\) and a query budget \(m\ll N_t\); the aim is to select \(\mathcal D_t^L\subset\mathcal D_t\) with \(|\mathcal D_t^L|=m\), keep the remainder unlabeled, and train a target model \(h_t(x)=g_t(f_t(x))\) that performs well on the target domain under the source-free constraint [2205.10711]. Subsequent formulations express the same objective as minimizing target risk with only source-pretrained parameters, unlabeled target data, and actively queried labels, sometimes in multiple rounds and sometimes in one shot [2407.18899], [2509.24896]. The topic now spans closed-set classification, segmentation, detection, multimodal supervision, and open-set medical image segmentation [2205.12840], [2407.02893], [2606.08749].

## 1. Definition and formal setting

ASFDA extends standard Source-Free Domain Adaptation (SFDA) by allowing a small number of actively chosen target labels. MHPL states this directly: ASFDA is a more practical scenario than SFDA because it permits actively selecting a few target data to be labeled by experts, while preserving the condition that source data cannot be accessed during adaptation [2205.10711]. LFTL describes the same setting as Source data-Free Active Domain Adaptation (SFADA): a source-pretrained model, an unlabeled target pool \(T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}\), a total annotation budget \(B\), and a target model \(M_t\) optimized to minimize
\[
R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]
\]
through iterative query-and-adaptation rounds [2407.18899].

A closely related formulation is given by DAM. There, a source-pretrained classifier \(S\) is trained over \(C\) classes, an unlabeled target set \(\mathcal T=\{x_j'\}_{j=1}^n\) is available, and a labeling budget \(B=\rho|\mathcal T|\) together with an active query strategy \(\pi\) selects a subset \(\mathcal T_q\subset\mathcal T\). The goal is to adapt a target model \(T\), initialized as \(T=S\), to minimize
\[
\mathcal R_T=\mathbb E_{(x',y')\sim\mathcal T}\big[\ell(T(x'),y')\big]
\]
using only \(\mathcal T\), the budgeted labels in \(\mathcal T_q\), and the source-model parameters [2509.24896].

Open-set variants generalize the label-space assumption. ASFOSDA defines source-free open-set domain adaptation for medical image segmentation by allowing unknown target classes \(C_u=C_t\setminus C_s\), so that \(C_t>C_s\). In that setting, the target task must handle both known source classes and private target-only structures, and negative transfer can occur if unknown classes are force-aligned to source features [2606.08749]. This establishes that ASFDA is not confined to the closed-set assumption used in early classification benchmarks.

## 2. Emergence of the paradigm and representative frameworks

MHPL is the first work that explicitly introduces the term “active source free domain adaptation (ASFDA)” and defines the core problem as active target selection under a source-free constraint [2205.10711]. Almost concurrently, SALAD formulates “Source-free Active Label-Agnostic Domain Adaptation” as a setting in which a pre-trained source-domain network must be adapted to a target domain with a small budget for annotation and a shift in the label space, again without access to source data [2205.12840]. These two 2022 works established two durable themes: active query design and adaptation under missing source data.

Later work broadened both methodology and scope. LFTL formulates SFADA as an iterative query-and-adaptation process and emphasizes “continuous performance improvements through the iterative query-and-adaptation process” [2407.18899]. ProULearn focuses on “sample exploration” through homogeneity propensity estimation, correlation-based uncertainty, and compact class distributions during adaptation [2501.13517]. DAM introduces a multimodal foundation model into SFADA and explicitly fuses human supervision with Vision-and-Language supervision through dual-focused active supervision and bidirectional distillation [2509.24896]. UGTST and ASFOSDA extend the setting to medical image segmentation, with ASFOSDA explicitly addressing open-set adaptation without source data [2407.02893], [2606.08749].

| Framework | Core mechanisms | Scope |
|---|---|---|
| MHPL | minimum happy points, neighbor ambient uncertainty, neighbor diversity relaxation, one-shot querying, neighbor focal loss | ASFDA classification [2205.10711] |
| SALAD | GATN and HAL | classification, segmentation and detection [2205.12840] |
| LFTL | Contrastive Active Sampling and Visual Persistence-guided Adaptation | iterative SFADA [2407.18899] |
| ProULearn | homogeneity propensity estimation, correlation index, central correlation loss | SFADA classification [2501.13517] |
| DAM | dual-focused active supervision, bidirectional distillation | multimodal SFADA [2509.24896] |
| UGTST / ASFOSDA | tiered self-training / Target-refined Self-training, uncertainty and discrepancy-based selection | medical segmentation, including open-set [2407.02893], [2606.08749] |

This progression suggests that ASFDA evolved from a narrowly defined classification scenario into a broader family of source-free active adaptation problems with differing assumptions about label space, supervision sources, and query schedule.

## 3. Query strategies and notions of informativeness

The earliest explicit ASFDA acquisition principle is MHPL’s notion of “minimum happy” points. A target point is an MH point if it is simultaneously neighbor-chaotic, individual-different, and target-like [2205.10711]. Neighbor-chaotic behavior is quantified by neighbor ambient uncertainty:
\[
\mathrm{NAU}(x)=\mathrm{NP}(x)\times \mathrm{NA}(x),
\]
where \(\mathrm{NP}(x)\) is neighbor entropy over clustering-based pseudo-labels and \(\mathrm{NA}(x)\) is average feature-space similarity to the \(q\) nearest neighbors. Diversity is imposed by Neighbor Diversity Relaxation, which greedily selects candidates while rejecting those whose nearest neighbor has already been selected. MHPL also adopts one-shot querying: all \(m\) labels are selected before any target fine-tuning [2205.10711].

LFTL replaces one-shot selection with an iterative criterion based on both uncertainty and persistence. Its Contrastive Active Sampling computes softmax probabilities at each round and forms a contrastively decoded log-probability. The Best-vs-Second Best margin
\[
u_{cm}(x)=\tilde p(y_a|x)-\tilde p(y_b|x)
\]
defines informativeness, with smaller values indicating more confusion, while a class-level intransferability term \(u_{ct}(c)\) down-weights classes that are already easy. The final score is
\[
u(x)=u_{cm}(x)+\lambda\,u_{ct}(y_a),
\]
and samples are queried in ascending order of \(u(x)\) [2407.18899]. LFTL explicitly defines “informative” samples as those on which the current model remains most uncertain and “persistently challenging” samples as those that remain uncertain even after previous rounds of adaptation [2407.18899].

ProULearn develops a different acquisition logic aimed at avoiding noisy outliers. Its homogeneity propensity estimation adapts the Isolation Forest idea: dense regions require more random splits and therefore longer average path-lengths. A correlation index is then computed between feature embeddings, local neighbor probabilities are averaged, local entropy \(E_i\) is calculated, and the final selection score is
\[
U_i=h(x_i)\times E_i.
\]
Selection ranks samples by \(U_i\) and excludes already-selected samples’ \(K\) neighbors to promote diversity [2501.13517].

Medical segmentation work emphasizes uncertainty decomposition and spatial redundancy. UGTST estimates global uncertainty per slice through test-time perturbations, pixel-wise entropy, histogram discretization, primary local peak filtering, and a diversity-aware redundancy filter based on K-means\({++}\) over encoder features [2407.02893]. ASFOSDA separates Class-aware Decomposed Uncertainty into aleatoric uncertainty and epistemic uncertainty, both estimated via test-time augmentation, then applies Class-agnostic Prototype Discrepancy, combining Cross-Domain Discrepancy and Self-Domain Discrepancy, to select the final annotation set [2606.08749].

A common misconception is that ASFDA acquisition is equivalent to plain uncertainty sampling. The literature does not support that reduction. MHPL combines uncertainty with neighbor affinity and diversity; LFTL adds persistence and class-level intransferability; ProULearn adds homogeneity propensity and correlation structure; UGTST adds redundancy filtering; and ASFOSDA adds prototype discrepancy in an open-set setting [2205.10711], [2407.18899], [2501.13517], [2407.02893], [2606.08749].

## 4. Adaptation objectives after querying

Acquisition is only one half of ASFDA; the other half is how the queried labels are exploited together with the unlabeled target set. SALAD frames the problem as a joint optimization of sample selection and knowledge transfer. Its Guided Attention Transfer Network (GATN) enables feature distillation from the pre-trained network to the target network, and its active learning function HAL mines target samples using transfer-ability and uncertainty criteria [2205.12840]. SALAD’s abstract emphasizes three consequences of this design: task-agnostic operation across classification, segmentation, and detection; handling of shifts in output label space from the pre-trained source network to the target domain; and adaptation without access to source data [2205.12840].

MHPL uses a neighbor focal loss to emphasize queried MH points during training. Labeled queried samples receive a cross-entropy term weighted by \(\alpha\,\mathrm{NP}(x)\), pseudo-labeled unlabeled samples receive a lower weight \(\beta\), and the final objective adds entropy minimization and class-balance regularization:
\[
L=L_{\mathrm{NF}}+L_{\mathrm{ent}}+L_{\mathrm{div}}.
\]
The intended effect is to focus more strongly on the target samples whose labels were actively purchased [2205.10711].

LFTL’s Visual Persistence-guided Adaptation operates after each query round. Newly labeled samples become anchors, their feature embeddings are maintained in a momentum-updated memory, unlabeled samples are softly assigned to anchors, and adaptation uses three losses: supervised cross-entropy on the labeled set, an anchor-based clustering loss \(\mathcal L_{vpa}\), and unlabeled-data entropy minimization \(\mathcal L_{ent}\). The joint objective is
\[
\mathcal L_{total}=\mathcal L_{ce}+\beta_1\mathcal L_{vpa}+\beta_2\mathcal L_{ent}.
\]
The paper stresses that VPA learns from features of actively selected anchors obtained from previous intermediate models, facilitating feature distribution alignment and active sample exploitation [2407.18899].

ProULearn adapts through pseudo-label refinement and compact-class learning. It initializes class centroids from current target predictions, assigns pseudo-labels using refined similarity scores \(z_{i,c}=C(f(x_i),o_c)\times h(x_i)\), and optimizes a full objective
\[
\mathcal L=\mathcal L_{wce}+\mathcal L_{im}+\mathcal L_{cc},
\]
where \(\mathcal L_{cc}\) is the central correlation loss that draws features toward their assigned class centroids [2501.13517]. The paper argues that this explicitly refines pseudo-labels and compacts class clusters around prototypes.

DAM broadens the supervision model itself. A frozen Vision-and-Language model \(V\), such as CLIP, supplies multimodal pseudo-label distributions through image and text encoders and learnable category prompts. On the labeled target subset \(\mathcal T_q\), DAM forms dual supervision through human supervision for the target model \(T\) and prompt-tuning supervision for \(V\). Fusion then occurs through bidirectional distillation: \(T\to V\) on top-confidence samples and queried labels, and \(V\to T\) on the target set with higher weights on queried labels. The target model objective combines distillation, active-label cross-entropy, entropy minimization, and a diversity regularizer, while the ViL model optimizes distillation and prompt supervision; at inference, only \(T\) is retained, so there are no additional FLOPs from \(V\) [2509.24896].

In medical segmentation, self-training is the dominant post-query adaptation strategy. UGTST uses a two-stage tiered self-training schedule: first, the source model is adapted on manually labeled slices together with source-model pseudo-labels for the remaining target data; then new pseudo-labels are generated with the augmentation ensemble, and the model is retrained on mixed true and pseudo labels [2407.02893]. ASFOSDA similarly performs supervised fine-tuning on queried labels, generates pseudo-labels for the remaining unlabeled set, and then minimizes a joint Dice-plus-cross-entropy objective over both labeled and pseudo-labeled subsets [2606.08749].

## 5. Task coverage, label-space assumptions, and domain-specific variants

ASFDA was first studied on standard visual classification benchmarks, including Office-Home, Office-31, and VisDA-2017 or VisDA-C [2205.10711], [2407.18899], [2501.13517]. Even within that initial regime, methodological assumptions diverged: MHPL adopts one-shot querying before fine-tuning, whereas LFTL explicitly iterates querying and adaptation across rounds [2205.10711], [2407.18899].

SALAD is notable for broadening the task definition beyond classification. Its experiments cover digits classification on MNIST, SVHN, and VISDA, synthetic-to-real image segmentation from GTA5 to CityScapes, and document layout detection from PubLayNet to DSSE [2205.12840]. It is also explicitly label-agnostic in the sense that it can handle shifts in output label space from the pre-trained source network to the target domain [2205.12840]. This suggests that active source-free adaptation is not intrinsically tied to a fixed shared label set.

Medical image segmentation introduced a different set of constraints. UGTST studies cross-center prostate MRI segmentation with a 2D U-Net source model, target domains from BIDMC and HK, and a one-round annotation budget of \(5\%\) of slices [2407.02893]. ASFOSDA generalizes the problem to open-set volumetric medical image segmentation, where target domains include private classes of new anatomical structures or pathologies not present in the source data. Its evaluation uses MSD Spleen \(\rightarrow\) FLARE, FLARE \(\rightarrow\) AMOS CT, and Abdominal MR \(\rightarrow\) AMOS MR, with both 3D U-Net and Swin UNETR backbones [2606.08749].

Multimodal variants add another dimension. DAM argues that prior SFADA methods rely exclusively on data-driven pseudo-labels and sparse human labels, while multimodal or external knowledge is not deeply fused. It therefore integrates multimodal supervision from a ViL model with scarce human annotations to form a dual supervisory signal [2509.24896]. This shifts ASFDA from purely model-internal uncertainty and clustering mechanisms toward supervision fusion with foundation models.

Another common misconception is that ASFDA is necessarily a closed-set classification problem with a conventional pseudo-labeling loop. The published record includes label-space shift in SALAD, open-set medical segmentation in ASFOSDA, and multimodal foundation-model supervision in DAM [2205.12840], [2606.08749], [2509.24896].

## 6. Empirical record, efficiency, and research directions

Quantitatively, early classification results established the basic usefulness of the paradigm. With a \(5\%\) label budget, MHPL reports average target accuracies of \(79.1\%\) on Office-Home, \(93.5\%\) on Office-31, and \(91.2\%\) on VisDA-2017, outperforming the listed SFDA, active learning, and active DA baselines in its evaluation protocol [2205.10711]. SALAD reports improvements of \(0.5\%-31.3\%\), across datasets and tasks, over prior adaptation methods that assume access to large amounts of annotated source data for adaptation [2205.12840].

LFTL emphasizes both accuracy and computational efficiency. On VisDA-C with \(1\%\) labels, it achieves \(87.4\%\), compared with SFUDA SOTA \(87.3\%\) for SHOT++ and ADA SOTA \(89.7\%\) for LADA, which uses source; at \(5\%\) labels it reaches \(92.8\%\), a \(+1.5\%\) gain over MHPL and \(+0.7\%\) over LADA. On Office-Home it rises from \(78.8\%\) at \(5\%\) labels to \(85.1\%\) at \(10\%\) labels, and on Office-31 from \(93.5\%\) to \(95.4\%\) over the same budgets. The same paper reports that adaptation on VisDA-C completes in approximately \(0.3\,\mathrm h\) versus \(5\text{–}8\,\mathrm h\) for prior ADA or SFUDA methods such as SHOT and LADA [2407.18899].

ProULearn reports further improvements on four benchmark datasets. At \(5\%\) labels, it reaches \(94.0\%\) on Office-31 versus \(92.9\%\) for MHPL, \(79.0\%\) on Office-Home versus \(78.4\%\), \(78.6\%\) on DomainNet-126 versus \(77.0\%\), and \(91.8\%\) on VisDA-2017 versus \(90.3\%\) [2501.13517]. Its ablations attribute gains to the central correlation loss, the HPE-based sample-selection mechanism, and the use of correlation rather than Euclidean or cosine distance in the distribution metric [2501.13517].

DAM reports that under a \(5\%\) labeling budget, CoreSet + DAM attains \(93.4\%\) on Office-31, \(83.9\%\) on Office-Home, and \(92.0\%\) on VisDA-C, while MHPL + DAM attains \(94.1\%\), \(82.0\%\), and \(92.0\%\), respectively. The paper summarizes these as consistent \(2\text{–}8\%\) absolute gains over strong SFADA baselines and states that DAM outperforms even methods that access source data or that use only single-path pseudo-labels [2509.24896].

In medical segmentation, UGTST reports a marked improvement with a mere \(5\%\) annotation budget, specifically an average Dice score enhancement of \(9.78\%\) and \(7.58\%\) in two target domains compared with state-of-the-art methods, reaching \(83.46\pm4.39\%\) Dice on BIDMC and \(81.17\pm7.65\%\) on HK, with HD\(_{95}\) values of \(5.16\pm1.73\) and \(3.37\pm1.15\), respectively [2407.02893]. ASFOSDA reports, with a 3D U-Net backbone and \(5\%\) budget, \(80.1\%\pm8.0\) DSC on MSD\(\rightarrow\)FLARE, \(78.7\%\pm8.9\) on FLARE\(\rightarrow\)AMOS CT, and \(71.5\%\pm8.4\) on MR\(\rightarrow\)AMOS MR, each with \(p<0.01\) versus all other SFDA and ASFDA methods in its comparison. The same work states that ASFOSDA recovers approximately \(84\%\text{–}85\%\) of fully supervised performance with only \(5\%\) labels and improves DSC by about \(6\%\text{–}7\%\) over the next best ASFDA method on Swin UNETR [2606.08749].

The main methodological tension in the literature is between one-shot querying and iterative query-and-adaptation. MHPL and ProULearn favor one-time selection, whereas LFTL and DAM explicitly reuse intermediate models during adaptation [2205.10711], [2501.13517], [2407.18899], [2509.24896]. Another tension concerns supervision sources: classical ASFDA relies on queried labels and pseudo-labels, while DAM argues for deep fusion of sparse human labels with multimodal priors [2509.24896]. Future directions identified within the literature include joint optimization of the active sampling strategy \(\pi\) within the DAM loop, extension to open-set or partial-label SFADA, exploration of other foundation models such as ALIGN or adapter-based architectures, extension of ProULearn’s propensity concept to text and speech, incorporation of active augmentation of queried examples, and end-to-end optimization of the tree ensemble and feature extractor [2509.24896], [2501.13517].

Taken together, the literature presents ASFDA as a family of source-free adaptation problems in which the marginal value of a small annotation budget is amplified by structured query criteria and adaptation objectives. The field began with minimum happy points and task-agnostic source-free active transfer, and it has expanded toward iterative persistence-aware querying, correlation-based sample exploration, multimodal bidirectional distillation, and open-set medical segmentation [2205.10711], [2205.12840], [2407.18899], [2501.13517], [2509.24896], [2407.02893], [2606.08749].

Source: https://www.emergentmind.com/topics/active-source-free-domain-adaptation-asfda