| When analyzing real-world data,a very important task is to find samples that are significantly different from other samples,which is anomaly detection.Nowadays,with the increasing intricacy of practical problem and the growth of data,new forms of abnormal data continue to appear,and abnormal data with known or unknown types are inevitable.Therefore,anomaly detection has occupied a very important position in modern data analysis.From the early outlier detection based on statistical methods to the current anomaly detection algorithm based on deep learning,many scholars have studied anomaly detection from different perspectives,but there are still many problems that need to be solved in anomaly detection.Based on the advanced Generative Adversarial Network(GAN),this dissertation solves three different problems from three different perspectives,and proposes three anomaly detection algorithms:(1)ORGAN-KDE: From the perspective of combining with traditional statistical method,by combining Orthogonal Generative Adversarial Network(O-GAN)with kernel density estimation method,a lightweight anomaly detection algorithm is proposed,which mainly solves the problems of complex network structure of the current anomaly detection algorithm based on deep generative models;(2)LCR-GAN: From the perspective of combining with other generative model,by combining GAN with Variational Auto-Encoder(VAE),an anomaly detection algorithm that can learn crucial representation of normal samples is proposed,which mainly solves the problem of good reconstruction effect of abnormal samples that may occur in anomaly detection phase based on reconstruction methods;(3)DAT-AD: From the perspective of combining with other adversarial idea,by combining the adversarial idea in GAN with the virtual adversarial training idea,an anomaly detection algorithm with dual adversarial training is proposed.It mainly solves the problem of improving the effectiveness of the algorithm while ignoring the robustness of the algorithm.This dissertation conducted a large number of experiments on structured datasets and image datasets,including performance comparison experiments,anti-noise experiments,ablation experiments,and so on.Through the analysis of the experimental results,it is found that the methods proposed in this dissertation can solve some existing problems of anomaly detection to a certain extent,and can achieve a good anomaly detection effect. |
