Anomaly Detection by Unsupervised Adversarial Generative Self-labelling Autoencoder

Abstract

Conventionally, autoencoder is applied to unsupervised anomaly detection but suffers from over-lapping distributions between reconstruction errors of normal and anomalous samples, which is also known as lack of learnability, as the autoencoder does not know the characteristic difference between normal and anomalous samples (autoencoder is trained on the normal samples, without any anomalous samples). Recently, the generative adversarial network is proposed, provided as another method instead of the autoencoder. However, the training process is time-consuming and unstable compared to the autoencoder. We propose an unsupervised adversarial generative self-labeling autoencoder (AGSA), optimizing the discriminator and autoencoder by adversarial training, and integrating the supervised learning into unsupervised learning: The discriminator evolves by including pseudo-labeled samples for training, then the autoencoder is required to decorate itself such that the reconstructed samples by the autoencoder score high in the discriminator, therefore resembling the normal samples and avoiding being similar to anomalous ones. In this case, reconstruction errors of anomalous samples are maximized and that of normal ones are minimized, compared to the previous autoencoder. Therefore, the overlapping part is lessened, and we are safe and confident to self-label more samples for training discriminators. An adversarial cycle is formed. AGSA replaces the conventional decision criteria using an anomaly score which highly relies on the prior knowledge of the dataset and is sensitive to the extreme and noisy samples of the dataset, with a well-trained discriminator which is less affected by the extreme and noisy samples. Transfer learning is used for constructing the discriminator for accelerating the training and avoid overfitting. AGSA has been applied to the clouds dataset and marble dataset, achieving satisfying results.