Research On One-class Data Anomaly Detection Algorithm Based On Deep Neural Network

The development and popularization of computer technology has brought great changes to the society,people are closely related to it in all aspects of production and life.Various information systems established by computer technology have also produced exponentially growing data.Learning these data through data mining technology can dig out the valuable information hidden behind it,so as to help people analyze things more deeply,predict their future development,and make decisions.As an active research direction in data mining technology,data anomaly detection plays an increasingly important role in cyberspace security,financial monitoring,disease monitoring and social security.The aim of data anomaly detection is to detect abnormal samples from normal data.Since anomalies occur with a small probability and the abnormal data is not easy to obtain,normal data and abnormal data are very unbalanced in quantity.Therefore,if a two-class model is used for anomaly detection,there will be problems of insufficient training data and poor generalization ability.To solve above problems,based on the deep neural network theory,combined with the concept of one-class classification,this paper carries out relevant research on one-class adversarial data anomaly detection algorithm.The main research achievements and innovations of this paper are presented as follows.(1)One-class adversarial anomaly detection nets with compact representationsThe one-class adversarial anomaly detection method is widely used because it solves the problems of poor generalization ability and unusable abnormal data in anomaly detection,but it is also limited by the instability of Generative Adversarial Network training and the non-global optimal problem of feature extraction.To this end,this paper proposes One-class Adversarial Anomaly Detection Nets with Compact Representations(CROCAN),containing a feature extraction stage and an anomaly discrimination stage.In the feature extraction stage,we use a feature extraction method based on compact-representation autoencoder transformed from the original encoder by adding compact loss in the latent space.The newly added loss function makes the feature extraction not only focus on the features that contribute to the reconstruction of samples,but also focus more on the robust features shared in normal data,which serves the goal of anomaly detection more.In the anomaly discrimination stage,we propose a multiple scale feature matching complementary generative adversarial network.This method adds multi-scale feature matching loss to the complementary generative adversarial network,so that the generated data can be fitted at each intermediate layer of the discriminator,reducing the phenomenon of gradient disappearance and mode collapse and resulting in more stable training.Finally,fake samples distributed at the margins are generated,and the discriminator is trained as an anomaly detector.(2)One-class adversarial anomaly detection nets with class specific representationsFeature extraction is a very important part of the classification algorithm.It maps samples from high-dimensional feature space to lowdimensional feature space for subsequent anomaly detection and determines the upper limit of detection results.However,in one-class classification problem,the lack of abnormal data makes it difficult for the extracted features to have a clear boundary between classes,limiting the performance improvement.To this end,this paper proposes a new oneclass classification model called One-class Adversarial Anomaly Detection Nets with Class Specific Representations(CS-OCAN).Firstly,a twoiteration framework is designed to make reasonable use of the reference data generated by multiple scale feature matching complementary GAN.Secondly,we propose a new feature extractor called class-specific autoencoder.It transforms the semi-supervised problem into a supervised problem and has an additional loss function in latent space that aims to maximize inter-class distances between two classes and minimize intraclass variances.The algorithm is divided into four stages.The first and second stages are the same as CR-OCAN proposed in the first innovation point,and they are used as preliminary feature extraction and reference data generation respectively.The third stage uses class-specific autoencoder to do the final feature extraction,and the fourth stage is the training of the multi-scale feature matching complementary GAN,and finally completes the training of the anomaly detector.