UniAd: Unified Adversarial Alignment for Unsupervised Cross-Domain Industrial Anomaly Detection

In practical applications, early Industrial Anomaly Detection (IAD) methods are typically trained on specific scenarios with abundant labeled samples. However, their performance degrades significantly when generalized to other scenarios with different data distributions–data from distinct domains often vary in feature distribution, data volume, anomaly types, and other aspects. Furthermore, annotating anomalous data is labor–intensive and costly, posing a major challenge to the practical deployment of IAD methods. To address these limitations, we propose a Unified Adversarial (UniAd) alignment model for IAD, which is the first unsupervised domain adaptation approach specifically designed for industrial anomaly detection. Specifically, UniAd adopts a global–local adversarial alignment mechanism: it globally mitigates inter–domain distribution divergence and locally enforces the semantic consistency of normal patterns across source and target domains. Subsequently, an autoencoder trained exclusively on normal samples from the source domain is employed to classify unlabeled target domain samples as normal or anomalous. Extensive experiments conducted on three public IAD benchmark datasets demonstrate that our proposed UniAd achieves state–of–the–art (SOTA) performance.