Balancing Feature Imitation and Knowledge Consistency for Improving Anomaly Detection

Unsupervised anomaly detection methods are essential for addressing the challenge of accessing anomalous data in real-world scenarios. Among these methods, student-teacher framework-based methods have garnered increasing attention due to their low latency and competitive performance. This framework leverages the knowledge disparities between the teacher and student networks, particularly their differences in feature extraction capabilities, for anomaly detection. In this paper, we propose not only to leverage the process of feature imitation for anomaly scoring but also to exploit the consistency among neighboring layer groups to enhance anomaly detection performance. Experiment shows that our model can significantly improve anomaly detection performance by incorporating a knowledge consistency term into the scoring process.