Learning to Generate Pseudo Anomalies

Marcella Astrid, Muhammad Zaigham Zaheer, Seung-Ik Lee

Keywords: anomaly detection, generative model, pseudo anomaly, autoencoder

TL;DR: We propose learning mechanism to generate pseudo anomalies for one-class classification in anomaly detection.

Abstract: Due to rare occurrence of anomalous events, anomaly detection is often seen as one-class classification (OCC) problem. In this setting, an autoencoder (AE) is typically trained to reconstruct using only normal data in order to learn normalcy representations. It is expected that, at test time, the AE can well reconstruct normal data while poorly reconstructing anomalous data. However, anomalous data is often well reconstructed as well. This phenomenon can be attributed to the fact that when training AE with only normal data, the boundary between normal and abnormal data is unknown, consequently resulting in a boundary that includes the abnormal data as well. To alleviate this problem, we utilize pseudo anomalies to limit the reconstruction capability of an AE. Without imposing strong inductive bias, pseudo anomalies are generated by adding noise to the normal data. Moreover, to improve the quality of pseudo anomalies, we propose a learning mechanism to generate noise by exploiting the aforementioned weakness of AE, i.e., reconstructing anomalies too well. Evaluations on Ped2, Avenue, ShanghaiTech, and CIFAR-10 datasets demonstrate the effectiveness of our approach in improving the discriminative capability of AEs for anomaly detection.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning