Introduction
Unsupervised Domain Adaptation (UDA) has traditionally focused on bridging the gap between a labeled source domain and an unlabeled target domain. However, real-world deployments often encounter "common corruptions"—such as motion blur, sensor noise, and weather effects—that are not present in the training data. This paper, presented at ICCV, introduces Distributionally and Discretely Adversarial Regularization (DDAR), a novel framework designed to enhance a model's Robustness against Common Corruption (RaCC) without requiring target labels.
The Challenge of Common Corruptions
Standard UDA methods assume target data is "clean," leading to significant performance drops when faced with environmental noise. While adversarial robustness (defending against malicious perturbations) is well-studied, "benign" real-world corruption in UDA remains under-explored. Existing solutions often suffer from a "robustness-accuracy trade-off," where gains in robustness come at the cost of accuracy on clean data. DDAR aims to break this trade-off by providing a "plug-in" regularization that improves robustness while maintaining standard performance.
Methods: The DDAR Framework
Overview of the DDAR framework showing the transition from clean source/target domains to corrupted target domains. Source, p. 4
DDAR addresses these issues through two primary mechanisms:
- Distributionally Adversarial Regularization: Instead of point-wise attacks, DDAR employs a min-max optimization based on the Wasserstein metric. It identifies a "worst-case" distribution shift in the latent space, forcing the model to generalize across a broad spectrum of potential corruptions. This approach treats adversarial data as a proxy; if a model can handle a distribution-level adversarial shift, it is better equipped for rain, fog, or sensor noise.
- Discretely Adversarial Regularization: To prevent overfitting to pixel-level noise, the authors utilize Image Discretization (ID). By leveraging a pre-trained VQGAN to map continuous pixels into a discrete visual vocabulary, the framework transforms adversarial perturbations into natural-looking "in-distribution" augmentations. This "visual word" approach weakens the attack just enough to become a helpful training signal rather than a confusing one.
Experimental Setup and Datasets
The framework was evaluated on three major benchmarks—Office-31, Office-Home, and VisDa-2017—augmented with 15 types of ImageNet-C corruptions (e.g., Gaussian noise, defocus blur, JPEG compression) across five severity levels. This rigorous testing ensures the model's performance is validated against a wide variety of real-world scenarios.
Key Results
Comparison of mean Corruption Error (mCE) across different levels of severity for VAT, AFD, and the proposed DDAR. Source, p. 6
DDAR demonstrates significant improvements in robustness. On Office-31, it reduced the mean Corruption Error (mCE) from 43.7% to 30.5%. Crucially, unlike many adversarial training methods, DDAR maintains or even improves standard accuracy on clean data. For instance, on the VisDa-2017 dataset, mCE improved from 65.8% to 55.6% while preserving competitive clean-data performance. Theoretical analysis further supports these results, showing that DDAR provides a tighter upper bound for target risk compared to previous methods.
Conclusion
DDAR offers a theoretically grounded and practically effective solution for deploying UDA models in unpredictable real-world environments. While it introduces some computational overhead due to the VQGAN and min-max optimization, the gains in reliability make it a compelling choice for safety-critical applications. Future work may explore applying this framework to more complex tasks like semantic segmentation or object detection.
Summary updated