Toward Adversarial Robustness in Unlabeled Target Domains

Area of Science:

• Deep Learning
• Machine Learning Security
• Computer Vision

Background:

• Adversarial training (AT) enhances deep learning model robustness against adversarial attacks.
• Existing AT methods fail when training/testing data distributions differ or target data is unlabeled.
• This limitation hinders model generalization and security in real-world, unlabeled scenarios.

Purpose of the Study:

• To address the challenge of adversarial training in unlabeled target domains.
• To propose a novel framework, Unsupervised Cross-domain Adversarial Training (UCAT), for robust model development.
• To enable knowledge transfer from labeled source domains to unlabeled target domains under adversarial conditions.

Main Methods:

• UCAT leverages labeled source domain knowledge to guide training.
• It utilizes high-quality pseudo-labels generated for unlabeled target domain data.
• Discriminative and robust anchor representations from the source domain are employed.

Main Results:

• Models trained with UCAT demonstrate both high accuracy and strong robustness.
• UCAT effectively prevents adversarial samples from misleading the training process in unlabeled target domains.
• Ablation studies confirm the effectiveness of UCAT's proposed components.

Conclusions:

• UCAT offers a robust solution for adversarial training in unlabeled target domains.
• The framework successfully bridges the gap between labeled source and unlabeled target domains.
• UCAT advances the field of robust deep learning by enabling training under distribution shifts and unlabeled data.