Progressive Alignment for Robust Domain Adaptation

ICLR 2026 Conference Submission16642 Authors

Keywords: Unsupervised Domain Adaptation, Adversarial Training, Adversarial Attack

Abstract: Unsupervised Domain Adaptation (UDA) has advanced knowledge transfer between labeled source and unlabeled target domains, yet existing methods fall short in real-world scenarios where adversarial attacks threaten model reliability. Robustness against such attacks is essential but remains critically underexplored in UDA. Existing methods often treat domain alignment and adversarial defense as separate steps, causing unstable training, noisy pseudo-labels, and incomplete feature alignment ultimately limiting their effectiveness. Addressing both domain shift and adversarial robustness simultaneously is vital for deploying trustworthy models in dynamic, adversarial environments. In this work, we propose a robust UDA method from the perspective of multi-source and multi-target domain adaptation, treating clean and adversarial samples across both source and target as distinct domains. We aim to align both clean and adversarial domains across source and target within the adaptation framework. Therefore, we use progressive domain alignment strategy that explicitly aligns clean target features with multi-source domains through classifier discrepancy minimization, and implicitly aligns adversarial target features by enforcing classifier agreement on pseudo-labels. We find that this strategy effectively handles both domain shift and adversarial perturbations, leading to improved generalization and robustness. We demonstrate the effectiveness of our approach through extensive experiments on four benchmark datasets, accompanied by component-wise ablations. Our method achieves standard accuracies of 62.0%, 88.4%, 82.5%, and 73.7% and the corresponding robust accuracies under PGD-20 attack with $\epsilon = 2/255$ are 49.4%, 78.3%, 77.3%, and 72.1% on the Office-Home, PACS, VisDA, and Digit benchmark datasets, respectively.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16642