Anomaly Detection With Divergence Regularized Autoencode

Anomaly detection has broad application prospects in fraud detection,data cleaning and predictive maintenance.In recent years,researchers have begun to apply deep learning to anomaly detection tasks and many breakthroughs have been made.Anomaly detection includes one-class classification and out-of-distribution detection.The difference between the above two directions is that one-class classification only uses single-class data(normal)to train the model,while the out-of-distribution detection training data set is multi-class.Nowadays,a method based on reconstruction error is commonly used to classify one class.The model demonstrates high reconstruction ability on normal data by training only on normal samples,but the reconstruction effect of abnormal data is very poor.However,due to the generalization of the model,reconstruction errors will fail.Besides,out-of-distribution detection often leads to over-fitting of the model due to the unreasonable structure of the auxiliary dataset.Thus,this thesis proposes two methods,namely,one-class classification based on divergence regularization and out-of-distribution detection based on disentangled auto-encoder to solve the above two problems.1)This thesis proposes a classification method based on divergence regularization and maximizing mutual information,which aims to improve the model’s ability to represent normal samples.This method first maximizes the mutual information between the input and the corresponding latent representation to make the model characterize the typical characteristics of normal samples,and then uses divergence regularization to limit the decision boundary of the latent space.Experimental results show that this method can effectively increase the accuracy of one-class classification.2)This thesis presents an out-of-distribution detection method based on the disentangled auto-encoder.We first use causal decomposition to design a disentangled auto-encoder,and then intervene on it to generate pseudo out-of-distributed data.Finally,we use normal data and generated data to train the model for uncertainty estimation.Experiments show that our method has a good performance improvement for out-of-distribution detection.