Research On Anomaly Detection Based On Batch Quadratic Programming Network

Anomaly detection involves machine learning,data mining,statistics,information theory and other related disciplines,and is widely used in intrusion detection,fraud detection,medical and health anomaly detection,network public opinion anomaly detection,industrial fault detection and other technical fields.In the anomaly detection,the existence of data noise or data errors determines the difficulty of anomaly detection based on data-driven method.Secondly,it is also difficult in anomaly detection to identify the real abnormal samples(abnormal exposure)from the samples containing noise.By suppressing or constraining samples with noise,the anomaly detection model will abandon such samples near the decision boundary,resulting that confusing the boundary between samples with noise and abnormal samples,and increasing the difficulty of anomaly exposure.Therefore,the influence of balancing noise suppression and anomaly exposure on the model is beneficial to improve the detection performance.This thesis proposes Batch Quadratic Programming network with maximum entropy constraint anomaly detection algorithm,including BQP network consists of feature extraction network and QP output layer,where feature extraction network maps batch samples to feature space,output batch feature vectors and QP output layer constructs SVDD dual problem of quadratic optimization constraint of batch training samples,outputs the optimal dual variable solution of this problem,and realizes SVDD hypersphere modeling of feature space of noise suppression.The uncertain sample set(including noisy samples and abnormal samples)was extracted from the batch feature vectors by the optimal dual variable,and the maximum entropy constraint loss function was used to enhance the role of anomaly exposure in the network optimization process and balance the influence of noise suppression and anomaly exposure.For some specific training data sets,such as the abnormal sample capture is relatively easy and constitutes a certain size of data set,the introduction of negative sample learning mechanism in model training can enhance the expression ability of the model.This thesis designed BQP network with semi-supervised learning method,and proposed Batch Quadratic Programming based on semisupervised learning network SSBQP(Semi-Supervised Batch Quadratic Programming,SSBQP)network.Before training,semi-supervised training data containing partial label information were prepared by triplet sampling method.During training,according to the label information of batch training samples,the adaptive SVDD quadratic constraint was proposed,and the constraint condition of SVDD problem was modified,so that SSBQP network could more reasonably match the label information of batch samples when processing different batches of samples.In the experiment,this thesis designed three control comparison experiments and set the baseline method of the control group to verify the detection performance of BQP network and SSBQP network.