Research On Anomaly Detection Method Based On Deep Learning

According to Hawkins’ definition of outliers: Anomalies are sample data that are so different from most observed data that it is suspected that they are generated by different mechanisms.The purpose of anomaly detection is to find outliers that are different from large-scale data through data mining,and to find the valuable knowledge behind the outliers.Anomaly detection has a wide range of applications,including but not limited to fraud detection,intrusion detection,environmental sanitation,image processing,medical processing,video network monitoring,trajectory anomaly detection and other fields.However,in the current big data environment,traditional anomaly detection models face the problems of high sparseness of high-dimensional data,slow calculation of massive data,and difficulty in defining anomalies in different scenarios.Many scholars have begun to shift the focus of anomaly detection research to more recent Years of development of hot deep learning.Based on the above background,in view of the limitations of traditional methods and some shortcomings of deep learning models,this paper proposes two improved deep anomaly detection methods.The main contents of the paper are as follows:(1)Proposed anomaly detection model based on improved adversarial autoencoder and ensemble learningFirstly,the traditional adversarial autoencoder is improved,and the discriminator is added to optimize the adversarial autoencoder.Second,the latent vectors are weighted and fused to make them more representative.Finally,we use the Ada Boost algorithm in ensemble learning to combine multiple decision trees into a strong classifier.In the experimental part,we conduct anomaly detection verification on two image datasets,MNIST and CIFAR-10.The comparison experiments with the current mainstream anomaly detection methods show that our proposed model has more advantages.(2)Proposed anomaly detection model based on shared autoencoderFirstly,the auto-encoder is improved,and additional encoders are added to form an encoding-decoding-re-encoding structure,and the two decoders share two encoders.Secondly,a two-stage training method with model matching is proposed,which combines the advantages of reconstruction training and adversarial training,while avoiding the shortcomings of generative adversarial networks and autoencoders.Finally,we still use the MNIST and CIFAR-10 image datasets for the verification of anomaly detection in the experimental part.Compared with the previous model and the latest anomaly detection model,it is not only effective but also time efficient.Finally,the paper is summarized and prospected.