Exploring Adversarially Robust Training for Unsupervised Domain Adaptation

Unsupervised Domain Adaptation (UDA) methods aim to transfer knowledge from a labeled source domain an unlabeled target domain. UDA has been extensively studied in the computer vision literature. Deep networks have shown be vulnerable adversarial attacks. However, very little focus is devoted improving robustness of deep models, causing serious concerns about model reliability. Adversarial Training (AT) considered most successful defense approach. Nevertheless, conventional AT requires ground-truth labels generate examples and train which limits its effectiveness In this paper, we explore robustify models: How enhance data via while learning domain-invariant features for UDA? To answer question, provide systematic study into multiple variants that can potentially applied UDA. Moreover, propose novel Adversarially Robust method accordingly, referred as ARTUDA. Extensive experiments on attacks benchmarks show ARTUDA consistently improves models. Code available at https://github.com/shaoyuanlo/ARTUDA .