MNGAN: Detecting Anomalies with Memorized Normal Patterns

Abstract

Anomaly detection is an significant problem in machine learning and has been well-studied in a wide range of applications. To model complex and high dimensional data distributions, existing methods, trained with an auto-encoder architecture either directly or indirectly, usually attempt to produce higher reconstruction error for anomalies than normal samples. However, lacking constrains on the latent representation of input data results in an unexpected performance that anomalies can be reconstructed well too, leading to the “missed alarm”. In this work, we propose to reconstruct input data with typical patterns of normal data learned through adversarial networks. Our approach, called MNGAN, which employs an encoder-decoder-encoder architecture with a memory network, learns to memorize prototypical patterns of normal data and simultaneously preserve details of data style for better reconstruction. In test phase, given a input data, the model will reconstruct it with the most relevant memory item, which indicates one normal pattern. Thus, reconstructions of anomalous data are similar to normal samples, resulting in effective detection for anomalies due to the high reconstruction error. Experiments over several benchmark datasets, from varying domains, shows that our proposed method outperforms previous state-of-the-art anomaly detection approaches.