Batch quadratic programming network with maximum entropy constraint for anomaly detection

Abstract

The difficulty of anomaly detection lies in balancing the impact of noise on the network (noise suppression) and distinguishing the real anomaly from the noise (abnormal exposure). So, a deep anomaly detector Batch Quadratic Programming (BQP) network with Maximum Entropy Constraint is proposed. It imposes quadratic programming constraints on Support Vector Data Description through the BQP output layer to achieve noise suppression. In BQP network processes’ batch data, Maximum Entropy Constraint is used to balance abnormal samples and noise. The experiment compared the shallow method with the currently popular deep method on MNIST and CIFAR-10 data sets and proved that the BQP network with Maximum Entropy Constraint has excellent performance.