Deep One-Class Classification Algorithm For Anomaly Detection

One-class classification(OCC)algorithm aims to build a classification model using only one class of target data.It plays an irreplaceable role in anomaly detection scenarios where nontarget data are insufficiently collected or defined.With the rapid growth of data in both quantity and dimension,the OCC anomaly detection algorithm based on the deep neural network is the current mainstream research direction,attracting a lot of attention.However,the current deep OCC anomaly detection algorithm still has the following two problems: 1)The lack of effective deep OCC loss functions specifically designed for anomaly detection,as existing OCC loss functions are mostly degenerated from multi-class loss functions and are not fully applicable to OCC anomaly detection problems.2)The lack of OCC anomaly detection loss functions in complex backgrounds,as in real-world applications,the collected data usually contains not only target data but also a small amount of undefined non-target data,and there is a lack of research on this issue.Therefore,this thesis conducts the following research work around the two above-mentioned issues:1.A OCC algorithm based on logarithmic barrier loss function(LBL)is designed for anomaly detection.Firstly,a strong inequality constraint is constructed for OCC,and then the logarithmic barrier function is used to smoothly approximate the inequality constraint,thus constructing the LBL.Specifically,LBL assigns larger gradients to samples near the hypersphere to accelerate the difficult samples’ convergence to the center of hypersphere,thereby obtaining a hypersphere with the minimum radius.In this way,the representation ability of the model for target data is improved,and the OCC anomaly detection performance is improved.In order to verify the performance of the proposed algorithm,comparative experiments with advanced algorithms are carried out on different network structures..2.A OCC loss function based on the Laplacian matrix and reciprocal distance constraints(LRL)is constructed for anomaly detection in complex scenarios.Firstly,a graph Laplacian regularization constraint is constructed for the target data based on the intra-class interaction distance,which guides data clustering through interaction transmission among intra-class samples,improving the robustness of model to noise.Secondly,a non-target data penalty item based on the reciprocal distance is designed to maximize distance between the non-target data and the center of hypersphere.Finally,the performance of the proposed algorithm is verified by comparison on many benchmark datasets.