| With the development of information technology and the popularization of artificial intelligence and big data technology,anomaly detection and analysis methods have attracted widespread attention in the academic community and have gradually become a research hotspot.However,in the collection and analysis of abnormal data,there is often a problem that the data gap between normal samples and abnormal samples is too large,and in some cases,the cost of collecting abnormal data is high or even difficult to collect abnormal samples.Therefore,how to achieve an efficient balance between the accuracy of anomaly detection data and the lack of sample data has become an important research direction in this field.On the one hand,this paper is based on the advantages of the one-class classifier technology,that is,the detection model is only trained with normal data,which can still maintain the performance comparable to the traditional anomaly detection method;on the other hand,the autoencoder neural network model is built based on the deep learning platform,and the model uses a nonlinear function to scale the dimension of the input samples,which has the advantage of being suitable for feature learning of various types of datasets.In this paper,we delve into autoencoders and their model variants and improve existing one-class classification anomaly detection methods from the perspectives of data reconstruction and data dimensionality reduction.The main work of this paper is as follows:(1)A reconstruction-based one-class classification anomaly detection method is proposed.Aiming at the problem that the traditional autoencoder needs to pre-set suitable parameters for different data sets to achieve the optimal detection effect.This paper first studies a data preprocessing strategy,which can enhance the characteristics of normal samples,which not only speeds up the convergence efficiency of the subsequent model,but also improves the detection accuracy of the model;Then,by studying the number of hidden layers,the number of layers,and the parameter update mechanism in the autoencoder,an adaptive encoder is constructed,and the adaptive encoder is trained by using the normal samples in the data set.Taking the reconstruction error of normal samples as anomaly threshold,anomaly detection is performed on the test set through the constructed adaptive encoder and anomaly threshold.The experimental results show that the reconstruction-based one-class anomaly detection method proposed in this paper achieves the optimal detection effect on four types of tabular datasets.(2)A hybrid anomaly detection method(SAAM-SVDD)with Self-Attention Autoencoder Model(SAAM)and Supported Data Domain Description(SVDD)is proposed.Aiming at the problem that traditional one-class anomaly detection methods have poor performance when dealing with high-dimensional data.In this paper,a two-stage detection method is constructed.First,Self-Attention autoencoder is used to extract abstract features from the preprocessed highdimensional image data,so as to obtain the low-dimensional representation of the original input sample and reduce the redundancy of the data;Then use the supported data domain description algorithm to divide the decision boundary of these low-dimensional features,and obtain a hypersphere that can wrap most of the samples.Anomaly detection is accomplished by extracting features from the test samples in the test set to see if the low-dimensional information falls inside the hypersphere.The experimental results show that the hybrid anomaly detection method proposed in this paper has improved detection accuracy compared with the comparison algorithms on three types of image datasets. |
