• 学术论文 •

### 基于深度对抗学习潜在表示分布的异常检测模型

1. 哈尔滨理工大学计算机科学与技术学院, 黑龙江 哈尔滨 150080
• 收稿日期:2020-09-03 修回日期:2021-02-20 出版日期:2021-07-25 发布日期:2021-08-11
• 作者简介:席　亮　男, 1983年5月出生于河北省邢台市. 现为哈尔滨理工大学副教授、硕士生导师. 主要研究方向为人工智能及其应用、网络与信息安全、深度学习等.E‑mail:xiliang@hrbust.edu.cn
刘　涵　男, 1996年10月出生于黑龙江省哈尔滨市.硕士研究生. 主要研究方向为人工智能及其应用，深度学习.E‑mail:liuhanharbin@163.com
樊好义　男, 1994年6月出生于河南省新乡市. 博士研究生，主要研究方向为网络嵌入、异常检测、时间序列信号分析、深度学习等.E‑mail:isfanhy@gmail.com
张凤斌　男, 1965年7月出生于黑龙江省哈尔滨市. 现为哈尔滨理工大学教授，博士生导师，CCF高级会员. 主要研究方向为网络与信息安全，人工智能与应用.E‑mail:zhangfb@hrbust.edu.cn
• 基金资助:
黑龙江省自然科学基金(F2018019)

### Deep Adversarial Learning Latent Representation Distribution Model for Anomaly Detection

Liang XI, Han LIU, Hao-yi FAN, Feng-bin ZHANG

1. School of Computer Science and Technology，Harbin University of Science and Technology，Harbin，Heilongjiang 150080，China
• Received:2020-09-03 Revised:2021-02-20 Online:2021-07-25 Published:2021-08-11

Abstract:

To solve the problems of the existing anomaly detection models, such as incoherent latent representation distribution under in high?dimensional and diverse(within each class) data background, the low accuracy of feature extraction when unbalanced data(normal data far outweighs abnormal data) is large, and the sensitivity of classifier’s hyperparameter, a deep adversarial learning latent representation distribution model for anomaly detection is proposed. Based on the regularization constraint, an improved autoencoder can map the original data feature space to a low?dimensional the latent feature space to get the reasonable latent representation distribution. On the premise of avoiding the problems of circulation inconsistent of reconstruction feature and unstable training, the multi?discriminator?based generative adversarial network can evaluate the latent representation probability distribution accurately, and to solve the hyperparameter sensitivity of one class classifier, so as to improve the overall performances of anomaly detection. Experimental results show that, compared with the up?to?date anomaly detection models based on machine learning and deep learning, the proposed model can obtain more coherent space distribution and ideal probability distribution of latent representation, is not sensitive to the hyperparameters of the single?class classifier, and effectively improve the detection performances under the application background with high?dimensional, diverse, unbalanced data.