Research On Anomaly Detection Method Based On Deep Generative Model

The semi-supervised anomaly detection model based on the deep generative model can learn the distribution of normal data and judge the data far from the distribution as anomalies.This type of model alleviates the problems of abnormal manifestations,so it has a wide range of application scenarios and application values.Variational AutoEncoder(VAE)and Adversarial AutoEncoder(AAE)are two important deep generative models used in semisupervised anomaly detection.However,the use of KL divergence(Kullback-Leibler Divergence)in VAE cannot measure the distance between the learned distribution of the model and the true distribution of the data.The anomaly detection model based on the AAE cannot explicitly learn the mutual information between normal data and its low-dimensional latent vector representation.Around these two issues,the specific work of this article is as follows:(1)A semi-supervised anomaly detection model based on Wasserstein AutoEncoder(WAE)is proposed.The model uses a training dataset composed of normal data for training,and uses the Maximum Mean Discrepancy(MMD)to represent the low-dimensional latent vector of normal data to constrain the latent space to follow the Gaussian prior distribution of the aggregation area to learn the complex distribution of normal data.Then use the hyperparameter search method in the verification dataset to determine the optimal anomaly detection threshold.Finally,according to the reconstruction probability of the data to be detected in the model,anomaly judgment is made.Experimental results on a variety of anomaly detection datasets show that compared with anomaly detection models based on Variational AutoEncoder,the F1 value of the model proposed in this paper are significantly improved.(2)A semi-supervised anomaly detection model(Infomax-AAE)which combines mutual information maximization target and Adversarial AutoEncoder is proposed.The model first trains the encoder and decoder with the reconstruction error minimization target;secondly,in the adversarial regularization stage,the distributed regularization constraint and the mutual information are maximized to train the encoder and discriminator as the training target,thereby reducing the low-dimensional of normal data the latent vector representation is constrained to a mixture Gaussian distribution and maximize the mutual information between the normal data and its low-dimensional latent vector representation;finally,the fully connected neural network is used to estimate the mutual information between the normal data and its low-dimensional latent vector representation.In the anomaly detection stage,combined with the reconstruction error of the data to be detected and the latent space divergence,a normalized anomaly score is calculated and anomaly judgment is performed.Experimental results on real financial datasets and two public anomaly detection datasets show that: adding mutual information to the training target of semi-supervised anomaly detection based on adversarial AutoEncoder to maximize the training target can significantly improve the anomaly detection performance of the model.