Improving Robustness of Unsupervised Domain Adaptation with Adversarial Networks

Abstract

Unsupervised Domain Adaptation (UDA) has been subject to comprehensive investigation and has achieved significant success in real-world scenarios by transferring information from labeled source domains to non-labeled target domains. Nonetheless, the vulnerability of UDA models to adversarial attacks remains a formidable challenge. While Adversarial Training (AT) is acknowledged as one of the most potent defense mechanisms, it cannot be directly applied to UDA settings. Furthermore, there have been few studies that explore the application of AT in UDA setting. In this paper, we strive to leverage Generative Adversarial Network (GAN) to generate adversarial examples for target data, subsequently incorporating them into AT. To generate high-quality adversarial examples and achieve better adversarial robustness of UDA models, we propose the AAT algorithm. We apply AAT to two different UDA algorithms and evaluate on three datasets. The results demonstrate that our model achieves improved adversarial robustness and a balance between accuracy on clean data and accuracy under adversarial conditions.