Source-Free Domain Adaptation

Source-free domain adaptation studies how to make a model trained in one setting work in a different setting when the original training data cannot be accessed, as often happens for privacy, storage, or ownership reasons. Common approaches adapt a pretrained model using unlabeled target data, relying on pseudo-labeling, consistency and entropy objectives, feature or prototype alignment, and teacher–student learning. Recent work also uses vision-language and other foundation models to provide semantic guidance, and extends adaptation to continual test-time use across tasks such as medical imaging, object detection, EEG, graphs, and time series. A central challenge is balancing adaptation to the new distribution against retaining useful source knowledge, since unreliable pseudo-labels can amplify errors or cause forgetting, especially under class imbalance, label shift, and severe real-world changes.