AED-PADA:Improving Generalizability of Adversarial Example Detection via Principal Adversarial Domain Adaptation

• DOI: 10.48550/arxiv.2404.12635
• Publication Date: 2024-04-19
• Authors: Heqi Peng; Yunhong Wang; Ruijie Yang; Beichen Li; Rui Wang; Yuanfang Guo
• Keywords: Domain Adaptation

Adversarial example detection, which can be conveniently applied in many scenarios, is important the area of adversarial defense. Unfortunately, existing detection methods suffer from poor generalization performance, because their training process usually relies on examples generated a single known attack and there exists large discrepancy between unseen testing examples. To address this issue, we propose novel method, named Example Detection via Principal Domain Adaptation (AED-PADA). Specifically, our approach identifies Domains (PADs), i.e., combination features different attacks, possesses coverage entire feature space. Then, pioneer to exploit multi-source domain adaptation with PADs as source domains. Experiments demonstrate superior ability proposed AED-PADA. Note that superiority particularly achieved challenging scenarios characterized by employing minimal magnitude constraint for perturbations.