基于记忆增强的对抗自编码器异常检测算法  

Anomaly detection algorithm based on memory⁃augmented adversarial autoencoder

作　　者：蔚焘 成卫青 WEI Tao;CHENG Weiqing(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学计算机学院、软件学院、网络空间安全学院,江苏南京210023

出　　处：《南京邮电大学学报（自然科学版）》2021年第6期84-94,共11页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61170322);江苏省研究生教育教学改革课题(JGZZ19_038)资助项目。

摘　　要：深度自编码器是异常检测领域中被广泛使用的深度学习模型。记忆增强的自编码器模型(Memory⁃augmented Autoencoder Model,MemAE)通过记忆增强模块解决传统自编码器泛化能力过强的问题,并取得了良好的效果。针对自编码器对于训练数据的正常模式提取能力有限这个问题,通过融合对抗自编码器(Adversarial Autoencoder,AAE)的对抗学习过程,提出基于记忆增强的对抗自编码器模型(Memory⁃augmented Adversarial Autoencoder Model,MemAAE)。相比于原模型,增加判别器模块,将自编码器与记忆模块视为生成器,使生成器输出的聚合后验分布与先验分布相匹配,弥补了自编码器能力不足的缺点,提高了重构的效果,并在一定程度上避免了可能出现的模式崩溃问题。在多个文本数据集和MNIST图像数据集上的实验结果表明,改进之后的模型与已有的异常检测模型相比,文中提出的模型MemAAE总体性能更优。The deep autoencoder is a deep learning model widely used in the field of anomaly detection.The memory⁃augmented autoencoder model(MemAE)solves the problem of excessive generalization ability of traditional autoencoder through memory⁃augmented module and achieves good effects.To overcome the problem that the autoencoder has limited ability to extract the normal mode of training data,a memory⁃augmented adversarial autoencoder model(MemAAE)is established based on the adversarial learning process of adversarial autoencoder.Compared with the original model,the discriminator module is added to the model,considering the autoencoder and the memory module as a generator,so that the aggregate posterior distribution of the generator output matches the prior distribution This compensates for the lack of autoencoder capabilities,thus improving the effect of reconstruction and avoiding the possible mode collapse to a certain extent.Experimental results on several text datasets and MNIST show that compared with the existing anomaly detection models,the model MemAAE achieves better performance on the whole.

关 键 词：异常检测 记忆网络 生成式对抗网络 对抗自编码器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]