Cloud-GAN: Cloud Generation Adversarial Networks for anomaly detection

Abnormal detection means identifying data that is different from the normal data. In recent work, there have been many methods using deep autoencoders or variational autoencoders to detect abnormal data, and good progress has been made. However, these methods often have the following problems: firstly, the value of low dimensional representations of different layers of potential space is different. And secondly, the effective information is not fully preserved, resulting in unsatisfactory reconstruction results. In this paper, we propose an anomaly detection method based on an adaptive cloud generation adversarial networks (Cloud-GAN), which uses three digital characteristics: expected value, entropy, and hyper entropy to represent the low dimensional representation of data in a more detailed manner in the generator. At the same time, the reconstruction loss weights of different layers are automatically learned. Then we input the reconstruction errors corresponding to the input and output data, as well as the low-dimensional representation of the last layer of the encoder, into the Gaussian kernel density model for density estimation. In a large number of experiments on six public benchmark datasets, we have verified that our method is equivalent to or superior compared to the state-of-the-art models. In particular, the effect is significant in high-dimensional dataset Arrhythmia with fewer training samples and the standard F1-score can be increased by up to 13% compared with the classical method LAKE.