Collective Anomaly Detection for Multivariate Data using Generative Adversarial Networks

Generative adversarial network (GAN) is used to model complex high-dimensional distributions of real-world scenarios. It has been applied to anomaly detection and making great achievements. However, most of the existing GAN-based anomaly detection methods cannot detect collective anomalies that change the behavior of multipoint data instances. Moreover, although many GAN-based methods for time-series anomaly detection have been proposed, there are few studies to handle collective anomalies in time-series data. Besides, there is still much room to improve the methods in terms of computational cost and the accuracy for detecting anomaly. We thus aim to propose a GAN-based method to detect multi-dimensional collective anomalies with high accuracy. To correctly detect collective anomalies, we especially introduce an encoder into a GAN-based anomaly detection method to obtain the latent states of the real data. We furthermore adopt a sequence to sequence technique to both encoder and generator, recurrent neural network, and fully connected neural network for the discriminator. We conducted experiments using two types of datasets: artificial and natural, and verified the effectiveness of our GAN model.