Dual-Student Knowledge Distillation-Based Anomaly Detection Algorithm

To address the diminished representational divergence on anomalous samples caused by identical teacher-student network architectures, this paper proposes DSAD (Dual-Student-based Anomaly Detection), an anomaly detection algorithm grounded in dual-student knowledge distillation. DSAD integrates two core components: (1) a dual-student network featuring structurally inverted student encoders and decoders, which breaks the conventional symmetry constraint in knowledge distillation to amplify teacher-student representational discrepancies for anomalies; and (2) an attention-based feature embedding module that effectively consolidates multi-scale features from student encoders. By incorporating attention mechanisms, this module mitigates information loss during multi-scale fusion, thereby delivering more compact embedded features to the student decoder.