PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey on Source-free Domain Adaptation

Zhiqi Yu, Jingjing Li, Zhekai Du, Lei Zhu, Heng Tao Shen

arXiv:2302.11803v1cs.LGcs.CVcs.MM

TL;DR

SFDA addresses domain adaptation when source data cannot be accessed because of privacy and security concerns, using a source-trained model and unlabeled target data. This survey organizes SFDA methods into a unified taxonomy, modularizes representative approaches, and compares them across three benchmarks, finding top methods within 3.6%, 2.7%, and 0.9% of target-supervision accuracy on the reported datasets.

  • Problem

    Conventional domain adaptation requires simultaneous access to source and target data, creating privacy and confidentiality concerns and motivating source-free adaptation.

  • Method

    The survey develops a unified data-centric and model-centric taxonomy, modularizes more than 30 representative methods, and compares their components and combinations across three classification benchmarks.

  • Results

    Top SFDA accuracies were within 3.6% on Office, 2.7% on Office-Home, and 0.9% on VisDA-C of target-supervision accuracy.

  • Takeaways & Limitations

    The survey provides a unified framework for comparing SFDA methods and identifies domain-based reconstruction as a regime extending UDA research to source-free settings.

  • Takeaways & Limitations

    Virtual-domain generation can be costly and may struggle with domain generalization when underlying data patterns are complex.

Abstract

from arXiv · show

Over the past decade, domain adaptation has become a widely studied branch of transfer learning that aims to improve performance on target domains by leveraging knowledge from the source domain. Conventional domain adaptation methods often assume access to both source and target domain data simultaneously, which may not be feasible in real-world scenarios due to privacy and confidentiality concerns. As a result, the research of Source-Free Domain Adaptation (SFDA) has drawn growing attention in recent years, which only utilizes the source-trained model and unlabeled target data to adapt to the target domain. Despite the rapid explosion of SFDA work, yet there has no timely and comprehensive survey in the field. To fill this gap, we provide a comprehensive survey of recent advances in SFDA and organize them into a unified categorization scheme based on the framework of transfer learning. Instead of presenting each approach independently, we modularize several components of each method to more clearly illustrate their relationships and mechanics in light of the composite properties of each method. Furthermore, we compare the results of more than 30 representative SFDA methods on three popular classification benchmarks, namely Office-31, Office-home, and VisDA, to explore the effectiveness of various technical routes and the combination effects among them. Additionally, we briefly introduce the applications of SFDA and related fields. Drawing from our analysis of the challenges facing SFDA, we offer some insights into future research directions and potential settings.

1 INTRODUCTION

SFDA addresses domain adaptation when source data cannot be accessed, using a source-trained model and unlabeled target data. This survey organizes SFDA methods into a unified, modular taxonomy and compares more than 30 representative works across three benchmarks.

  • Problem setting: SFDA differs from UDA because it adapts a source-trained model with unlabeled target data when source data are unavailable.Conventional UDA methods use both source and target data, making them unsuitable for settings without source-domain access.
  • Taxonomy: The survey divides SFDA into data-based and model-based directions while also discussing domain-based reconstruction as an extension of UDA.Domain-based reconstruction reconstructs or subdivides domains to compensate for missing source data.
  • Method components: Model-centric SFDA methods commonly use self-training, including pseudo-labeling, entropy minimization, and contrastive learning.Pseudo-labeling methods focus on prototype generation, pseudo-label assignment, and pseudo-label filtering to obtain reliable target supervision.
  • Contribution: The survey proposes a comprehensive taxonomy from data-centric and model-centric perspectives to summarize SFDA methods and applications.It is presented as a response to the lack of a timely comprehensive survey of the field.
  • Evaluation and analysis: The survey modularizes more than 30 representative SFDA works and compares their results on Office-31, Office-home, and VisDA.The component-level comparison is intended to expose relationships among technical routes and their combination effects.

2 OVERVIEW

The survey focuses primarily on the single-source, closed-set SFDA setting, where source data are available only during pre-training and adaptation uses unlabeled target data. It formalizes the domain and task, while noting extensions to broader domain and adaptation settings.

  • Preliminaries: The section introduces notation and preliminary knowledge for SFDA before discussing its different settings.These preliminaries are summarized in a notation table and related definitions.
  • Domain definition: In the typical closed-set SFDA setting, source and target label spaces are the same, although their marginal distributions differ.The source domain is complete during pre-training but is not available during adaptation.
  • Task definition: SFDA adaptation begins with supervised source-model pre-training and then uses unlabeled target data after source-domain data become unavailable.The objective is to obtain accurate target predictions from the adapted model.
  • Scope and variants: The survey mainly covers single-source closed-set SFDA, while outlining partial, open-set, multi-source, and multi-task variants.SFDA can also be combined with test-time, federated, and active domain adaptation settings.

3 DATA-BASED METHODS

Data-based SFDA methods compensate for unavailable source data by reconstructing virtual or perturbed domains and exploiting target-domain structure. The survey organizes these approaches around virtual domain generation, intra-domain adversarial alignment, perturbed domain supervision, and related consistency-based design choices.

  • Overview: Data-based SFDA methods reconstruct a domain or divide target data to compensate for missing source data and extend UDA techniques.The survey identifies virtual domain generation, intra-domain adversarial alignment, and perturbed domain supervision as principal reconstruction directions.
  • Virtual Domain Generation: Virtual domains are generated through adversarial models or Gaussian-distribution modeling, using information implicit in the pretrained source model.Adversarial generation may condition a generator on predefined labels, while Gaussian approaches model source-like distributions from classifier or batch-normalization statistics.
  • Virtual Domain Generation: Consistency losses constrain semantic agreement across generated and target domains, with designs including semantic, domain-discriminator, deterministic, weight, clustering, and contrastive constraints.Examples include 3C-GAN, SDDA, and CPGA, which emphasize semantic similarity, domain-invariant features, or representative class prototypes.
  • Virtual Domain Generation: Non-generative virtual-source construction selects reliable target samples, but selected samples may comprise only one-tenth of the target domain, limiting distributional coverage.Mixup, uncertainty-aware selection, and continued transfer of high-confidence samples are used to enlarge or improve the virtual source domain.
  • Perturbed Domain Supervision: Perturbed-domain methods assume shared domain-invariant features and design target-related perturbations that improve generalization without destroying original data or model properties.Neighborhood clustering additionally exploits target-data structure by enforcing consistency among neighboring samples.

4 MODEL-BASED METHODS

Model-based SFDA methods adapt selected model components using unlabeled target predictions, with self-training as the dominant paradigm. The survey organizes self-training into pseudo-labeling, entropy minimization, and contrastive learning, emphasizing modular combinations and label-noise control.

  • Model-based methods separate the model into sub-modules and adjust selected parameters for domain adaptation.
  • Self-training is the dominant model-based paradigm, using target predictions to refine the model through pseudo-labeling, entropy minimization, and contrastive learning.Auxiliary modules are often used to improve robustness and preserve source-domain knowledge.
  • 4.1.1 Pseudo Labeling: Pseudo-labeling generates class prototypes, assigns labels by prototype similarity or distance, and filters noisy labels during target adaptation.Without prototypes, labels can instead be assigned directly from the model’s most confident output category.
  • 4.1.1 Pseudo Labeling: Class prototypes may be selected from high-confidence samples, classifier weights, or multiple prototypes intended to represent each target-domain class distribution.The survey notes that simple prototype-selection strategies may not produce representative prototypes.
  • 4.1.1 Pseudo Labeling: False-label filtering rejects unreliable pseudo-labels through rules or additional network mechanisms, while noisy-label learning uses strategies such as adaptive thresholding.Filtering improves pseudo-label accuracy but can introduce additional resource overhead.
  • 4.1.2 Entropy Minimization: Entropy minimization constrains model predictions directly on target data and is readily combined with pseudo-labeling and other adaptation methods.Batch entropy maximization can be added to maintain prediction diversity because entropy minimization alone may bias outputs toward one class.
  • 4.1.2 Entropy Minimization: The survey describes information maximization as combining entropy minimization with batch-level diversity constraints, while noting limited innovation beyond such extensions.Future work may examine how to constrain target outputs at individual and batch levels while preserving reasonable diversity.
  • 4.1.3 Contrastive Learning: Contrastive learning adapts representations by gathering positive pairs and separating negative pairs, but source-free settings must construct queries that represent unavailable source information.The survey distinguishes memory-bank-, encoder-, and mini-batch-based contrastive approaches.

5 COMPARISON

The survey compares modularized SFDA methods across Office-31, Office-Home, and VisDA-C, finding that leading methods approach target-supervised accuracy while different technical routes show distinct strengths.

  • 5.1 Comparison Datasets: The survey mainly reports results on Office-31, Office-Home, and VisDA-C because Digits offers limited improvement and DomainNet is difficult.The datasets include Digits, VisDA-C, Office-31, Office-Home, and DomainNet, but the latter three benchmarks are emphasized.
  • 5 COMPARISON: The comparison modularizes SFDA methods into four classes and nine modules to analyze component effectiveness and combinations.The four classes are Domain-based Reconstruction, Image-based Information Extraction, Self-Training, and Self-Attention.
  • 5.2 Comparison Results: 3.6%, 2.7%, and 0.9% are the respective gaps between the best SFDA accuracies and target-supervised accuracy on Office-31, Office-Home, and VisDA-C.The reported best methods are Kundu et al. + SHOT++ on Office-31, TransDA on Office-Home, and BMD + SHOT++ on VisDA-C.
  • 5.2 Comparison Results: Self-training is the most popular SFDA research line, with pseudo-labeling commonly combined with entropy minimization or contrastive learning.Pseudo-labeling remains vulnerable to noisy labels and error accumulation, motivating filtering and implicit regularization approaches.
  • 5.2 Comparison Results: 90.7% on Office-31 and 88.7% on VisDA are reported for Kundu et al. and BMD, respectively, using SHOT++ as a baseline.The survey characterizes pseudo-labeling as versatile, effective, and a strong baseline for improvement.
  • 5.2 Comparison Results: Domain-based reconstruction is commonly used, while virtual domain generation combines closely with self-training and reaches 90.7% on Office-31.Intra-domain adversarial alignment methods were earlier approaches, but selected source-similar target samples may be noisy and unrepresentative.
  • 5.2 Comparison Results: Neighborhood clustering outperforms image style transformation in the reported comparisons, with SCLM leading Cl→Ar and Cl→Rw on Office-Home.AaD achieves 88% on VisDA using one loss function and is presented as a simple, strong baseline.
  • 5.2 Comparison Results: 79.3% on Office-Home is achieved by transformer-based TransDA, 4.8% above the second-highest method.The survey describes self-attention as promising, although it is less used in image classification than in semantic segmentation.

6 APPLICATION

SFDA extends domain adaptation’s privacy-oriented setting across computer vision and NLP applications. Most work remains in computer vision, while NLP applications are still relatively limited.

  • 6 APPLICATION: SFDA preserves source-domain privacy while reducing the cost of labeling target-domain data, overlapping substantially with unsupervised domain adaptation applications.The paper identifies computer vision and natural language processing as the main application fields.
  • 6 APPLICATION: Image classification is the dominant SFDA task and extends to image dehazing, cross-scene hyperspectral classification, and blind image quality assessment.The paper also discusses semantic segmentation and other computer-vision applications.
  • 6 APPLICATION: NLP applications remain relatively limited, including continuous learning, pretrained-model generalization, negation detection, and time-expression recognition.The paper cites SemEval 2021 Task 10 and extensions of self-training, active learning, and data augmentation baselines.
  • 6 APPLICATION: Source-free domain generalization, source-free zero-shot domain adaptation, and test-time domain adaptation are presented as related settings.These settings address previously unseen domains, reduced streaming-data requirements, or test-time constraints.

7 FUTURE RESEARCH DIRECTION

The survey identifies limits in SFDA’s current research scope, theoretical foundations, applications, datasets, and evaluation practices. It highlights broader settings, stronger theory, more diverse tasks and data, and comprehensive evaluation as directions for extending the field.

  • SFDA research remains concentrated in limited and homogeneous problem settings, with classification tasks dominating other computer vision tasks.
  • Most existing methods focus on pseudo-labeling, while perturbed domain supervision, neighborhood clustering, transformer-based self-attention, contrastive learning, and image style translation remain less explored or relatively simple.
  • Current SFDA theory is often method-specific because source-target domain divergence is difficult to measure without source data, motivating universally applicable theoretical support.
  • SFDA has few video-analysis methods, and applications established in unsupervised adaptation for NLP, time-series analysis, recommendation systems, and geosciences have not yet broadly extended to source-free settings.
  • Most evaluation datasets use balanced, clean data, while some benchmarks are too easy—for example, Office-31 W to D already reaches 100% accuracy—underscoring the need for more challenging datasets and broader evaluation.A comprehensive evaluation should include model robustness and overhead.
  • The survey mainly covers single-source closed-set SFDA, leaving partial-, open-set, multi-source, multi-task, test-time, federated, and active settings as additional combinations or extensions.

8 CONCLUSION

The survey reviews source-free adaptation, where source data are unavailable, through a unified framework and comparative analysis. It examines more than 30 representative methods across three major datasets, modularizes their components, and proposes future research directions.

  • The survey presents a comprehensive overview and unified categorization framework for existing source-free adaptation methods.
  • It analyzes more than 30 representative SFDA methods on Office-31, Office-home, and VisDA, while modularizing methods to facilitate comparison.
  • The paper suggests potential research directions based on its analysis and the current state of SFDA research.
Loading 2302.11803v1…