Introduction
Traditional Unsupervised Domain Adaptation (UDA) relies on simultaneous access to a labeled source dataset and an unlabeled target dataset. However, real-world constraints such as data privacy (e.g., medical records) and storage limitations (e.g., massive datasets on edge devices) often necessitate Source-free Domain Adaptation (SFDA). In SFDA, a model must adapt to a new target domain using only a pre-trained source model, without access to the original source data.
The primary challenge in SFDA is the "frozen" nature of the source model. When target data deviates significantly from the source distribution—forming a "source-dissimilar" set—accuracy typically drops. Existing methods often struggle to re-align these dissimilar samples without the source data as a reference.
Methodology
The proposed Adaptive Adversarial Network ($A^2Net$) addresses these challenges through three integrated components:
- Adaptive Adversarial Inference: $A^2Net$ employs a dual-classifier strategy. It utilizes a frozen source classifier ($C_s$) and a learnable target classifier ($C_t$). Target samples are adaptively categorized into "source-similar" and "source-dissimilar" sets. An adversarial game is then played: the feature generator attempts to align "hard" (dissimilar) samples with "easy" (similar) ones, while the classifiers attempt to distinguish between them. This discrepancy-based adversarial signal effectively aligns the target distribution without needing source domain labels.
- Contrastive Category-wise Matching (CCM): To enhance intra-class compactness, $A^2Net$ performs contrastive learning on target image pairs. It uses a dynamic "easy-to-hard" curriculum, where similarity thresholds are adjusted over time to include increasingly ambiguous pairs as training progresses.
- Self-Supervised Rotation: A rotation prediction task (0°, 90°, 180°, 270°) is incorporated as a pretext task. This forces the feature extractor to capture robust geometric and structural information from the target images, improving feature representation.
Key Results
$A^2Net$ was evaluated on standard benchmarks including Office-31, Office-Home, and VisDA. It consistently achieved state-of-the-art performance among source-free methods and even rivaled "source-needed" techniques.
On the Office-31 dataset, $A^2Net$ achieved an average accuracy of 90.1%, surpassing the popular SHOT method (88.6%). On the more complex Office-Home dataset, it reached 72.8% average accuracy. Most notably, on the large-scale VisDA synthetic-to-real challenge, it scored 84.3%, a substantial improvement over the source-only baseline of 52.4%. These results suggest that modeling the internal structure of the target domain can be as effective as, or even superior to, explicit source-target alignment.
Conclusion
$A^2Net$ demonstrates that effective domain adaptation is possible without source data by leveraging adversarial classifier discrepancy and self-supervised structural learning. While it relies on a reasonably competent initial source model and involves iterative optimization, its performance across diverse benchmarks establishes it as a robust solution for privacy-preserving and resource-constrained transfer learning. Future work may explore more sophisticated self-supervised tasks to further enhance feature robustness.
Summary updated