Adaptive Boundary‐Guided Learning and Counterfactual Feature Shuffling for Anomaly Detection

Industrial anomaly detection relies on training with normal data alone, since anomalous samples are not available during the training phase. Conventional methods often demonstrate inaccurate boundary estimation while depending on unrealistic synthetic anomalies. To overcome these challenges, we propose a novel method that generates virtual anomalies directly from normal features through gradient‐based feature shifting and counterfactual feature shuffling, thereby facilitating accurate boundary formation in feature space. These virtual anomalies are used to train a discriminator that effectively differentiates between normal and anomalous representations, thus enhancing the accuracy of anomaly map generation. To evaluate the proposed method, experiments are performed on four publicly accessible datasets, MVTec AD, VisA, MPDD, and WFDD, using image‐level and pixel‐level AUROC metrics. The results indicate that the proposed method improves the detection performance and anomaly localisation compared to existing methods.