Abstract
Unsupervised multi-source domain adaptation (UMDA) transfers knowledge from multiple labeled source domains to an unlabeled target domain. Traditional UMDA assumes direct access to source data, raising significant privacy concerns. This paper introduces Trust center sample-based Source-free Domain Adaptation (TSDA), a method that operates without source data by leveraging pre-trained source models. TSDA identifies high-confidence "trust center samples" in the target domain and employs a self-supervised pseudo-labeling strategy. By adaptively weighting source-specific target models, TSDA achieves state-of-the-art performance on benchmark datasets and demonstrates superior robustness against adversarial perturbations.
Introduction
Unsupervised Multi-source Domain Adaptation (UMDA) is critical for deploying models across diverse environments. However, the requirement for concurrent access to source and target data is often impractical due to data privacy regulations, massive data volumes, or proprietary restrictions. Source-free UMDA addresses these challenges by adapting pre-trained source models to a target domain without ever seeing the original source data. The primary hurdle is aligning distributions and correcting domain shifts in the absence of source samples.
Methodology: The TSDA Framework
The proposed TSDA method follows a multi-stage pipeline to bridge the domain gap without source data.
1. Source Model Acquisition and Initialization
Models are first pre-trained on multiple distinct source domains using standard cross-entropy loss. These models serve as the starting point for adaptation, providing initial feature extractors ($F^i$) and classifiers ($C^i$).
2. Trust Center Sample Selection
To handle the unlabeled target data, TSDA uses an entropy function to measure prediction uncertainty. Samples with the lowest entropy—representing the highest model confidence—are designated as "trust centers" for each class. These centers act as reliable anchors for the target domain's feature space.
3. Self-Supervised Pseudo-Labeling
Instead of using raw model predictions which can be noisy, TSDA assigns pseudo-labels to target samples based on their cosine similarity to the identified trust centers. This stabilizes the learning process and provides a cleaner supervisory signal.
4. Adaptive Weight Integration
Not all source domains are equally relevant to the target. TSDA employs a multi-layer perceptron (MLP) to calculate contribution weights for each source model. This allows the system to prioritize knowledge from source domains that are most similar to the target.
5. Robust Optimization
The target model is optimized using a composite loss function:
- Information Maximization ($L_{ent} + L_{div}$): Minimizes uncertainty while ensuring class diversity in predictions.
- Classification Loss ($L_{cls}$): Trains the model using the refined pseudo-labels.
Experimental Evaluation
TSDA was evaluated on several standard benchmarks: Office-Home, Office-Caltech10, and Office31. Additionally, the researchers tested robustness using adversarial samples generated via Projected Gradient Descent (PGD).
Key Results
- State-of-the-Art Accuracy: TSDA achieved an average accuracy of 98.0% on Office-Caltech10, surpassing both source-free and source-dependent baselines.
- Adversarial Robustness: TSDA showed remarkable resilience. Under 100% target data perturbation on the Office-Home dataset, TSDA maintained 54.1% accuracy, outperforming methods like ATEN (53.5%) and DECISION (51.8%).
- Effective Source Weighting: The adaptive mechanism correctly identified domain similarities, such as the high relevance of "Webcam" data when the target is "DSLR".
Discussion and Insights
A significant finding is that TSDA often outperforms traditional UMDA methods that have full access to source data. This suggests that the source-free constraint may act as a regularizer, encouraging the model to learn more robust, domain-invariant features directly from the target distribution. By filtering for "trust centers" first, the model avoids the error propagation common in standard pseudo-labeling approaches.
However, the current reliance on a single trust center per class may be a limitation for datasets with high intra-class variance. Future work will explore multi-center selection strategies and extending the framework to natural language processing tasks.
Conclusion
TSDA provides a powerful, privacy-preserving solution for multi-source domain adaptation. By combining entropy-based sample selection with adaptive model weighting, it achieves superior performance and robustness without the need for sensitive source data.
Summary updated