«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年07期

页码:

55-62

栏目:

网络空间安全

出版日期:

2025-07-10

文章信息/Info

Title:

Anomaly Network Intrusion Detection Method Based on Federated Learning with Global Knowledge Distillation

文章编号:

1673-629X(2025)07-0055-08

作者:

石迎澳1; 李润知1; 2; 姬怡1

1. 郑州大学 网络空间安全学院,河南 郑州 450001;

2. 郑州大学 网络管理中心,河南 郑州 450001

Author(s):

SHI Ying-ao1; LI Run-zhi1; 2; JI Yi1

1. School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450001,China;

2. Network Management Center,Zhengzhou University,Zhengzhou 450001,China

关键词:

入侵检测系统; 联邦学习; 数据隐私; 知识蒸馏; 数据异构

Keywords:

intrusion detection system; federated learning; data privacy; knowledge distillation; data heterogeneity

分类号:

TP399

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0087

摘要:

在网络入侵检测领域,联邦学习(FL)作为一种保护数据隐私的分布式处理方法受到了广泛关注。 然而,由于参与节点间的数据异构性,传统的联邦学习方法在联合训练过程中往往难以实现高性能。 为了解决这一问题,该文提出了一种改进的联邦学习方法,即基于联邦全局知识蒸馏的异常网络入侵检测方法(FLGKD-ANIDS)。 该方法在中央服务器中设置了一个缓冲区,用于缓存客户端上传的多轮模型参数。 进一步地,这些缓存的参数被用于生成包含多轮全局知识的教师模型参数,指导客户端侧的知识蒸馏过程。 这一机制使得客户端能够在注入全局知识特征的情况下训练本地数据。实验在两个公开数据集 UNSW-NB15 和 CIC-IDS2017 上进行,结果显示 FLGKD-ANIDS 在各种数据异构场景下显著提升了模型性能,其性能更接近于集中训练模型的水平。

Abstract:

In the field of network intrusion detection,Federated Learning (FL) has gained significant attention as a distributed processing method that protects data privacy. However,due to data heterogeneity among participating nodes,traditional FL methods often fail to achieve high performance during joint training. To address this issue,we propose an improved federated learning with global knowledge distillation (FLGKD-ANIDS). It sets up a buffer in the central server to cache multiple rounds of model parameters uploaded by clients.Furthermore,these cached parameters are used to generate teacher model parameters containing multi-round global knowledge,which guide the knowledge distillation process in the client side. This mechanism allows clients to train the local data by injecting global knowledge feature. We conducted experiments on two public available datasets,UNSW-NB15 and CIC-IDS2017. The results show that FLGKD-ANIDS significantly improves model performance across various data heterogeneity scenarios compared to existing federated learning methods,and its performance approaches that of centrally trained models.

常用功能

更新日期/Last Update: 2025-07-10