Anomaly Detection With Generative Adversarial Networks In Time Series

With rapid development of computer networks and the Internet of Things,massive amounts of time-series data are continuously being collected in various industries.Therefore,anomaly detection is crucial in many industries,such as intrusion detection in cyber security,fault diagnosis in industry and health monitoring.Time series contain a large number of complex spatio-temporal semantic features,e.g.cycles and trends.However,it is difficult to perform complex annotations or even to obtain anomaly samples in reality.This makes it challenging for anomaly detection in time series.To address these problems,this thesis investigates the following research contents.Since the one-class classification method usually handles the anomaly detection task as a representation learning problem or a feature modeling problem,it cannot solve the problems of insufficient training samples and inadequate feature representation.This thesis proposes a temporal anomaly detection method,called GSAD,based on generative adversarial networks and the one-class classification method SVDD.First,the auto-encoder is used to compress the samples to obtain the feature vectors,while a normal distribution constraint is applied to the feature vector,and two discriminators are constructed to classify and learn the input temporal data and the generated features respectively,constituting a cyclic consistency-based generative adversarial network.The adversarial learning method is used to solve the problems of insufficient sample and inadequate feature representation in the training dataset,with the purpose of increasing robustness in the model.Second,SVDD is used to perform in one-class classification on the generated feature vectors,so that the features generated by the encoder can have a compact feature representation.The benefits are that the model convergence can be accelerated,and the entropy can be reduced in the system.The traditional deep learning-based SVDD method requires pre-training the autoencoder as well as computing the hypersphere center.In comparison,our no longer requires a twostage training model,and it is capable to perform joint training of the GAN and SVDD and it dose not need to compute the hypersphere center separately.The experiment results demonstrate that the validity of the model can be verified on the MSL and SMAP datasets to assist anomaly detection for spacecraft telemetry signal.To address the problem that some labeled anomalous samples are discarded and not fully utilized in unsupervised learning,this thesis proposes a semi-supervised anomaly detection method,called SGSAD,based on unsupervised GSAD,in which a small amount of labeled samples,especially anomalous samples,are collected for semisupervised learning in unsupervised learning of a large amount of unlabeled samples,which can significantly improve the anomaly detection performance.The purpose is to design new loss functions for generative adversarial networks and support vector description methods for semi-supervised learning of anomaly detection,respectively.In our proposed method,the feature representation of normal samples has a low entropy,while the feature representation of anomalous samples has a high entropy.The experiment demonstrats that by adding a small amount of labeled data into a large amount of unlabeled dataset,the semi-supervised learning can improve the F1 metrics by about 4% to 6%,compared with the GSAD model.