Research On Anomaly Detection Of High-dimensional Data Based On Ensemble Generative Adversarial Networks

Anomalies are objects that are significantly different from others in a dataset,which often contain some important information.Anomaly detection of high-dimensional sparse data is still a challenge due to the curse of dimensionality and the sparsity of the data.It is valuable to explore anomaly detection approaches for high-dimensional data.The idea of data generation is introduced to address such problems in the thesis.(1)An unsupervised anomaly detection approach combining Generative Adversarial Network(GAN)with ensemble learning,named GAN＿Ensemble,is proposed.Owing to the ability of the generator of a GAN to simulate the distribution of real data,a large volume of latent data can be generated to avoid the data space to be too sparse.Moreover,in model training,multiple pairs of generators and discriminators are fully connected with each other.Thus,the mismatch of generators and discriminators will enhance the generative model to learn more complex data distributions,and to improve anomaly detection effect finally.Experiments on public datasets show that the index AUC can be improved by 7% averagely compared to traditional GAN-based anomaly detection approaches.At the same time,there is also an increase of 7.5% to 21.8% on AUC compared with the classical anomaly detection approaches.(2)Considering that the proposed approach GAN＿Ensemble may drop into overfitting and has high time consumption,it is optimized further and an unsupervised anomaly detection approach DGANs based on selective ensemble Generative Adversarial Networks is proposed.DGANs removes connections between generators and discriminators with a specific probability during the training phase.This makes the connections between generators and discriminators sparser.For anomaly detection,the well trained discriminators are integrated selectively based on dynamical voting weights adjusting.This can not only avoid the model falling into overfitting,but also reduce the time cost and improve the detection.Experiments on public datasets show that DGANs can improve the average accuracy by 4.63% compared with GAN＿Ensemble.Moreover,DGANs also show advantages on recall rate and F1 score compared to classical anomaly detection approaches.