Research On Deep Adversarial Learning Latent Representation Distribution Model For Anomaly Detection

Anomaly detection aims to identify and detect data that is different from some features of normal data.Because the existing anomaly detection models cannot obtain a reasonable latent representation distribution of data in the context of high-dimensional and diverse samples(various within the class),When there are many unbalanced data(normal data is much larger than abnormal data),the quality of feature extraction is low,and the classifier hyperparameters are sensitive.Therefore,we propose an anomaly detection model based on deep adversarial learning latent representation distribution.In order to obtain a reasonable latent representation space distribution to solve the problems of high dimensionality and hyperparameters sensitivity,we use regularization-based AE(RAE)to model the data.Firstly,map the data from the original feature space to the latent feature space to form a low-dimensional latent representation;and then,the improved regularization term increases the distance between latent representations and shortens the distance within the classes,so that the latent representations are concentrated in the latent feature space and realize a more rationalized spatial distribution of latent representations.In order to accurately estimate the probability distribution of latent representations to solve the problem of low quality of feature extraction when there are diverse samples(various within the class)and more imbalanced data,we designs a deep adversarial learning latent representation distribution model for anomaly detection(DALR),under a reasonable latent representation space distribution,using "latent representations,original features and reconstruction features" as input for adversarial learning,on the basis of effectively avoiding inconsistency reconstruction feature cycles and unstable training of the generative adversarial network,Effectively estimate the probability distribution of the latent representations;then,the obtained probability distribution of the latent representation as the input of the one-class classifier,thereby effectively improving the overall performance of anomaly detection.The experimental results show that compared with state-of-the-art anomaly detection method based on machine learning and deep learning,DALR can obtain a more reasonable spatial distribution and effectively estimate it in the conte xt of high-dimensional,diverse samples,and more imbalanced data applications.The proposed model can avoid the hyperparameter sensitivity of one-class classifiers and improve the detection performance effectively.