基于GAN的异常检测研究综述  

Review of Research on Anomaly Detection Based on GAN

作　　者：樊富有 代洋[3] 张淋 FAN Fuyou;DAI Yang;ZHANG Lin(Intelligent Terminal Key Laboratory of Sichuan Province,Yibin,Sichuan 644000,China;Network and Library and Information Center,Yibin University,Yibin,Sichuan 644000,China;Electronic Information Engineering,China West Normal University,Nanchong,Sichuan 637000,China)

机构地区：[1]智能终端四川省重点实验室,四川宜宾644000 [2]宜宾学院网络与图书情报信息中心,四川宜宾644000 [3]西华师范大学电子信息工程学院,四川南充637000

出　　处：《宜宾学院学报》2023年第6期1-10,共10页Journal of Yibin University

基　　金：智能终端四川省重点实验室开放课题项目(SCITLAB-0019);网络与数据安全四川省重点实验室开放课题项目(NDSZD201603);四川省教育厅重点项目(17ZA0452)。

摘　　要：深度学习中生成式对抗网络(GAN)具有强大拟合训练数据分布能力,在应用到异常检测领域时可有效准确识别异常图像.针对异常检测领域中传统有监督学习算法存在大量已知标记样本训练的局限性,以无监督学习GAN的异常检测模型为研究对象,阐明生成对抗网络的基本原理、网络结构及相关理论,详细介绍了近年来十种典型的基于GAN的异常检测模型,经过比较各衍生模型的异同,总结出各自的优势、局限性和应用场景,通过分析GAN在异常检测领域研究中所面临的问题及挑战,展望了未来的研究方向主要是解决模型的稳定性、计算效率、生成样本的精度、异常区域定位、异常评价机制等问题.The biggest advantage of Generative Adversarial Network(GAN)in deep learning is its strong ability to fit the distri⁃bution of training data,which can be applied to anomaly detection field effectively and accurately identify abnormal images.Aim⁃ing at the limitation of traditional supervised learning algorithms in the field of anomaly detection with a large number of known marker samples,the unsupervised learning GAN anomaly detection model was taken as the research object.The basic principles,network structure and related theories of generative confrontation networks were classified and ten typical anomaly detection models based on GAN in recent years were introduced in detail.By comparing the similarities and differences of the derived mod⁃els,the advantages,limitations and application scenarios were summarized.By analyzing the problems and challenges faced by GAN in the field of anomaly detection,the future research direction is mainly to solve the problems of model stability,computa⁃tional efficiency,sample generation accuracy,anomaly location,anomaly evaluation mechanism and so on.

关 键 词：深度学习 生成式对抗网络 异常检测 无监督学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]