Collective Anomaly Detection using Generative Adversarial Networks

Abstract

Generative adversarial network (GAN) is now being applied to anomaly detection. However, the existing approaches to GAN-based anomaly detection cannot detect collective anomalies that change the behavior of some data instances because they deal with individual data instances. This study aims to determine how collective anomalies that are commonly associated with time-series data can be detected using GAN models. We developed a GAN model for time-series data by adopting a decoder side of sequence to sequence (seq2seq) to a generator, an encoder side of seq2seq to an encoder, recurrent neural networks and fully connected neural network to a discriminator. We conducted several experiments on datasets, regarded as anomaly datasets, that we generated by swapping data instances at different time points. The results suggest that our GAN model can compete effectively with existing approaches for detecting collective anomalies.