Semi-Supervised Anomaly Detection withContrastive Regularization

International audienceDeep anomaly detection has recently seen significantdevelopments to provide robust and efficient classifiers using onlya few anomalous samples. Many of those models consist in a firstisolated step of representation learning. However, in its currentform the learned representation does not encode the semantics ofnormal sample and anomalies. Indeed during the first step thesemodels will not utilize the available normal/anomaly labels, harmingthe downstream anomaly detection classifier performances.In the light of this limitation, we introduce a new deepanomaly detector enforcing an anomaly distance constraint onthe norm of the representations while using contrastive learningon the direction of the features. This allows it to learn representationswell-suited to anomaly detection while avoiding anyrepresentation collapse. Moreover, we introduce two strategies ofanomaly enriching to improve the robustness of any distance basedanomaly detector. Our model highly improves the state-of the-art performances on a wide array of anomaly types with upto 74% error relative improvement on object anomalie