As one of the main tasks of data analysis,anomaly detection is a subject of extensive research in data science.The one class classification method is one of the main methods in anomaly detection,which is aimed at the classification problem that the dataset contains only one class of samples.As the data size increases,traditional one class anomaly detection methods such as one class support vector machines and kernel density estimation often perform poorly in high-dimensional data scenarios due to their weak computational scalability and the curse of dimensionality.To be effective,these traditional methods usually require extensive feature engineering,so that many irrelevant features create noise in the input data,masking real anomalies.In recent years,many deep anomaly detection methods have been proposed,and the effective one is the deep support vector data description method.But in the implementation,this method will have serious problems such as "sphere collapse" not solved.In addition,the weakness of deep learning that is susceptible to adversarial examples will also have an impact on anomaly detection results.Therefore,this paper conducts in-depth research on the above-mentioned issues,and the main research contents and achievements include:(1)Construct a deep support vector data description method with a small number of outliers.This method builds a hypersphere around normal data to separate normal data and abnormal data,and adds a small number of abnormal samples to refine the hypersphere classification boundary,so that the model more robust,while solving the problem of "sphere collapse" in deep support vector data description method.Experiments are conducted on the MNIST,CIFAR10 and Fashion_MNIST public datasets,and the results show that,compared with the original Deep SVDD method,the AUC value on the MNIST dataset has an average increase of 0.1%.On the CIFAR10 dataset,the AUC value has increased by an average of 5%.On the Fashion_MNIST dataset,the AUC value is improved by an average of 2.6%.(2)Construct a deep support vector data description method for adversarial training: this method uses the idea of adversarial training to find adversarial sample points around normal data,and puts the adversarial sample points and normal data into a deep anomaly detection model for training,so that the model can resist adversarial attack,and improve the classification ability as well as abnormal detection ability of the model.This method solves the problem of difficulty in obtaining abnormal data,and because the adversarial samples are used as abnormal points,it also solves the problem of "sphere collapse" in the description of deep support vector data.Experiments on the Fashion_MNIST dataset show that compared with the original Deep SVDD method,the AUC value is increased by 4%on average.Compared with the Deep SVDD method with a small number of outliers,the AUC value is increased by an average of 2%. |