Source-Free Domain Adaptation

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2792 papers960 conceptsfrozen 2026-10-04

Traditional Unsupervised Domain Adaptation (UDA) assumes a luxury that real-world practitioners rarely have: simultaneous access to both a labeled source dataset and an unlabeled target dataset. In practice, data privacy regulations like GDPR, intellectual property concerns, and the sheer logistical burden of moving terabytes of data often mean that once a model is trained, the original data is deleted or locked away.

Source-Free Domain Adaptation (SFDA) addresses this constraint. It asks: How can we adapt a pre-trained model to a new environment using only the model's weights and unlabeled target samples?

The Core Challenge: Bridging a One-Sided Gap

In standard UDA, we align domains by minimizing the distance between source and target feature distributions. Without source data, the "source distribution" becomes a ghost. We have the decision boundaries (the classifier) and the feature extractor, but we don't know exactly where the source data points lived in the latent space.

The intuition behind SFDA is that the pre-trained model itself contains a "fingerprint" of the source domain. Researchers have developed three primary strategies to recover or substitute this missing information.

Strategy 1: Feature Statistics as Source Proxies

The most direct way to "remember" the source domain is to look at the internal statistics of the model. Marvin Klingner and colleagues demonstrated in Unsupervised BatchNorm Adaptation (UBNA) that Batch Normalization (BN) layers are particularly informative. Since BN layers store the running mean and variance of source features, we can adapt a model simply by updating these statistics with target data, effectively recalibrating the model's internal "normal" to the new environment.

Taking this further, Cian Eastwood and Ian Mason proposed Bottom-Up Feature Restoration (BUFR). Instead of just updating statistics, BUFR stores lightweight "fingerprints" of feature distributions and realigns them to handle measurement shifts, such as changes in camera hardware, without needing the raw images.

Strategy 2: Generative Reconstruction and Prototypes

If we can't see the source data, can we imagine it? Some methods attempt to reconstruct "avatar" versions of the source data. Yifan Zhang and Zhen Qiu developed CPGA, which mines the pre-trained model to generate feature prototypes for each class. These prototypes serve as anchors; the model then uses contrastive learning to pull unlabeled target samples toward their corresponding class anchors.

Similarly, Yunzhong Hou and Liang Zheng explored Source Free Domain Adaptation with Image Translation, where target images are translated into the "style" of the source domain by matching their feature statistics to the pre-trained model's BN layers. This allows the model to "see" the target data through a source-colored lens.

Strategy 3: Robust Pseudo-Labeling

The most common approach is self-training: using the source model to predict labels for the target data (pseudo-labels) and then fine-tuning the model on those predictions. However, because of the domain shift, these pseudo-labels are often noisy.

To solve this, researchers focus on label refinement:

  • Uncertainty and Consensus: Mattia Litrico proposed Guiding Pseudo-labels with Uncertainty Estimation, which uses neighbor consensus to filter out unreliable predictions.
  • Negative Learning: Pietro Morerio and Waqar Ahmed introduced AdaPLR, which employs "Negative Learning"—teaching the model what a sample is not rather than what it is—to mitigate the impact of incorrect pseudo-labels.
  • Hypothesis Disparity: Qicheng Lao and Mohammad Havaei developed HDMI, which uses an ensemble of models and penalizes them when they disagree on target data, forcing the adaptation to be more stable.

Beyond Single-Source Adaptation

Recent work has expanded SFDA to more complex scenarios. Zongyao Li and colleagues addressed the Union-Set Multi-source problem, where multiple pre-trained models (each covering different classes) must be merged into a single target model. Meanwhile, Maohao Shen and Gregory Wornell provided a theoretical foundation for this in On Balancing Bias and Variance in Unsupervised Multi-Source-Free Domain Adaptation, moving the field from heuristic tricks to rigorous information-theoretic frameworks.

By treating the pre-trained model as a compressed, privacy-preserving representation of the source domain, SFDA enables AI deployment in sensitive fields like medical imaging (Yunxiang Li, Source-free Domain Adaptation for Brain Skull Stripping) and autonomous driving (Divya Kothandaraman, SAfE), where data can never leave its original silo.

Papers in this version

Showing 60 of 792 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 1138 papers
  2. Version 1 · 2026-10-04 · 792 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_source_free_domain_adaptation_04a3d0e1,
  title = {source-free domain adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/source-free-domain-adaptation/rdr_04a3d0e169314468b17c950d04d83ce1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_source_free_domain_adaptation_04a3d0e1_v1,
  title = {Source-Free Domain Adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/source-free-domain-adaptation/rdr_04a3d0e169314468b17c950d04d83ce1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}