Anomaly Detection Algorithm Based On Memory Model

Anomaly detection is an important and challenging task,which requires the proposed anomaly detection model to be able to identify anomalies from vast amounts of data.From traditional machine learning algorithms to recent deep neural networks,anomaly detection tasks have been a hot research topic.Besides,the research for anomaly detection is a broad application prospect.In the field of intelligent security,it can be used to detect abnormal events in videos? in network security,it can be used to monitor network attacks and resist illegal attacks? in the field of intelligent manufacturing,it can be used to detect defects in industrial products.An inherent problem of anomaly detection is that anomalous samples are difficult to collect and label due to their diversity and scarcity,thus lacking supervised information that can be effectively exploited to optimize the anomaly detection model.Therefore,thesis mainly studies the unsupervised anomaly detection algorithm based on the memory model,which uses only normal data for training to obtain the ability to detect anomalies.With the help of the deep learning method and the cognitive memory model,thesis conducts a series of related researches on visual anomaly detection,intrusion detection,and industrial anomaly detection.The research contents and main contributions are as follows:1.Aiming at the shortcomings of visual anomaly detection algorithms,thesis pro-poses a novel network based on memory model to solve the visual anomaly detection task.The model is based on the structure of Autoencoder,and proposes to store the latent space features of normal images with the help of the memory model so that abnormal images will be reconstructed into normal images,so as to increase the reconstruction difference of abnormal images.Moreover,thesis proposes to combine the memory model with the density estimation network and use the density estimation network to learn the probabil-ity distribution of normal image data to detect abnormal images.Finally,thesis proposes a two-stage training method to accelerate the convergence of the model and improve its performance of the model.2.Aiming at the shortcomings of intrusion detection algorithms,thesis proposes a novel autoencoder based on memory model for Internet of Things(Io T)intrusion detec-tion.The model abandons the supervised learning method and adopts the unsupervised learning method of the autoencoder,which effectively solves the problems of unbalanced data set categories and the difficult detection of unknown anomalies.In thesis,the network structure data is preprocessed and input into the autoencoder based on a convolutional neu-ral network.At the same time,thesis proposes to use the memory model to store the latent space features of normal requests and uses the feature reconstruction loss and feature spar-sity loss to constrain the memory model,which enhances the representation ability of the memory model and the diversity of memory units.3.Aiming at the shortcomings of industrial anomaly detection algorithms,thesis proposes an innovative student-teacher network based on memory-split model for indus-trial anomaly detection.The model distills the teacher network's feature representation of normal data to the student network with a memory-split model and uses the difference in feature representation of the two networks for image regions as an evaluation criterion for abnormality.Thesis proposes to use a memory-split model to store normal features,thereby expanding the difference of feature expression for abnormal regions.In addition,the memory-split model proposes to store the feature map in different regions to improve the expressiveness of the spatial position and eliminate the insensitivity to the spatial po-sition of the image content.