信息网络安全 ›› 2025, Vol. 25 ›› Issue (3): 364-391.doi: 10.3969/j.issn.1671-1122.2025.03.002

• 综述论文 • 上一篇    下一篇

基于深度学习的时序数据异常检测研究综述

陈红松1,2(), 刘新蕊1, 陶子美1, 王志恒1   

  1. 1.北京科技大学计算机与通信工程学院,北京 100083
    2.材料领域知识工程北京市重点实验室,北京 100083
  • 收稿日期:2025-01-30 出版日期:2025-03-10 发布日期:2025-03-26
  • 通讯作者: 陈红松 E-mail:chenhs@ustb.edu.cn
  • 作者简介:陈红松(1977—),男,北京,教授,博士,CCF高级会员,主要研究方向为网络空间安全、人工智能与大数据|刘新蕊(1999—),女,河南,硕士研究生,主要研究方向为大数据时序异常检测模型|陶子美(2000—),女,河南,硕士研究生,主要研究方向为人工智能时序数据异常检测算法|王志恒(2000—),男,山东,硕士研究生,主要研究方向为网络空间安全中的人工智能模型优化
  • 基金资助:
    国家语委科研项目(YB145-110);国家重点研发计划(2023YFC3303800);北京市昌平区“科技副总”专项(KW202404006021)

A Survey of Anomaly Detection Model for Time Series Data Based on Deep Learning

CHEN Hongsong1,2(), LIU Xinrui1, TAO Zimei1, WANG Zhiheng1   

  1. 1. School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China
    2. Beijing Key Laboratory of Knowledge Engineering for Materials Science, Beijing 100083, China
  • Received:2025-01-30 Online:2025-03-10 Published:2025-03-26
  • Contact: CHEN Hongsong E-mail:chenhs@ustb.edu.cn

摘要:

时序数据异常检测是数据挖掘及网络安全领域的重要研究课题。文章以时序数据异常检测技术为研究对象,运用文献调研与比较分析方法,深入探讨了深度学习模型在该领域的应用及其研究进展。文章首先介绍了深度时序数据异常检测的定义与应用;其次,提出了深度时序数据异常检测面临的9个问题与挑战,并将时序数据异常分为10类,枚举了16种典型的时序数据异常检测数据集,其中包括5种社交网络舆情安全领域相关数据集;再次,文章将深度时序数据异常检测模型进行分类研究,分析总结了近50个相关模型,其中包括基于半监督增量学习的社交网络不良信息发布者异常检测,进一步地,文章依据深度学习模型的学习模式将模型划分为基于重构、基于预测、基于重构与预测融合3种类型,并对这些模型的优缺点及应用场景进行综合分析;最后,文章从8个方面展望了深度时序异常检测技术的未来研究方向,分析了每个方向的潜在研究价值及技术瓶颈。

关键词: 深度学习, 时序数据, 异常检测, 模型分类, 社交网络

Abstract:

Anomaly detection for time series data is an important area of data mining and network security research. This paper focuses on time series anomaly detection techniques, employing literature survey and comparison analysis to thoroughly examine the applications and research progress of deep learning models in this domain. The research firstly introduced the definition and applications of deep time series anomaly detection, followed by an identification of the nine key challenges faced in this area. Time series anomalies were categorized into ten types, and sixteen typical anomaly detection datasets were enumerated, including five datasets related to social network public opinion security. Deep time series anomaly detection models were classified, the paper categorized and summarized nearly fifty relevant models, including those based on semi-supervised incremental learning for detecting abnormal information disseminators in social networks. Furthermore, the research classified these models into three categories according to their learning modes: reconstruction-based, prediction-based, and a fusion model, their advantages, disadvantages and applications were compared. Finally, the research outlined future research directions for deep time series anomaly detection in eight key areas, providing comprehensive perspects on potential advancements in the fields, potential values and technological bottlenecks were analyzed.

Key words: deep learning, time series data, anomaly detection, model classification, social network

中图分类号: