Kernel Fence GAN: Unsupervised Anomaly Detection Model Based on Kernel Function

Data imbalance and difficult construction of high-dimensional data cause the low accuracy of anomaly detection. This paper proposes an unsupervised anomaly detection model Kernel Fence GAN to increase the accuracy. In this adversarial generative network, the generator's loss function consists of three items: the encirclement loss limits the generated samples to lie at the decision boundary of normal samples, the central dispersion loss maximizes the range of the decision boundary, and the enhanced dispersion loss uses kernel function to nonlinearly map the generated samples to RKHS space, which can maximize sample differences and avoid outliers. The purpose of training the discriminator is to accurately identify whether the input sample is a normal sample in the original dataset or an abnormal sample generated by the generator. In this paper, validation experiments are performed on two datasets: on two-dimensional synthetic dataset, the experiment result shows that the model has faster convergence rate than Fence GAN. On the KDD99 dataset, compared with traditional and advanced deep learning methods, both precision and recall have been correspondingly improved.