基于对比学习和关系蒸馏的增量式入侵检测方法
CSTR:
作者:
作者单位:

(天津大学 电气自动化与信息工程学院,天津 300072)

作者简介:

金志刚(1972—),男,教授,博士生导师

通讯作者:

金志刚,zgjin@tju.edu.cn

中图分类号:

TP393.08

基金项目:

国家自然科学基金(52171337)


An incremental intrusion detection method based on contrastive learning and relational distillation
Author:
Affiliation:

(School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China)

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为解决入侵检测系统在学习新类攻击样本时存在的灾难性遗忘、分类偏向新类、旧类泛化能力弱等问题,提出基于对比学习和关系蒸馏的增量式入侵检测方法(incremental intrusion detection method based on contrastive learning and relational distillation,CLRD-IIDM)。首先,提出对比学习增强方法,构建具有特征提取层、投影分支和分类分支的双分支网络结构,并提出旧类补偿对比损失与之配合以增强对旧类样本的记忆;其次,提出新旧类关系蒸馏方法,提取旧模型中的新旧类相似关系并构造类关系标签,利用这些标签对新模型施加正则化约束,缓解模型对新类的分类偏向问题;最后,设计均衡偏差示例存储器,存储能够有效且全面表征旧类特征的样本以保持模型对旧类的泛化能力。在UNSW-NB15数据集上的实验结果表明,应用CLRD-IIDM的多个神经网络模型的最终准确率优于基线模型2.11%~9.78%;对比实验和消融实验结果验证了本文方法的有效性和泛化性。CLRD-IIDM能较好地适应复杂多变的现实网络环境,具有一定的应用价值。

    Abstract:

    To solve the problems of catastrophic forgetting, classification bias towards new classes, and weak generalization ability of old classes in intrusion detection systems when learning new attack samples, an incremental intrusion detection method based on contrastive learning and relational distillation(CLRD-IIDM) is proposed. Firstly, a contrastive learning enhancement method is introduced, which constructs a dual branch network structure with the feature extraction layer, projection branch, and classification branch. The old class compensation contrastive loss is employed to cooperate with it to enhance the memory of old class samples; Secondly, the distillation method for the relationship between old and new classes is proposed, which extracts the similarity relationship between old and new classes in the old model and constructs class relationship labels. These labels are used to apply regularization constraints to the new model to alleviate the problem of classification bias towards new classes; Finally, the balanced bias example memory is designed to store samples that can effectively and comprehensively characterize the features of the old class, which maintains the models generalization ability to the old class. The experimental results on the UNSW-NB15 dataset show that the final accuracy of multiple neural network models using CLRD-IIDM is 2.11% to 9.78% better than the baseline model. Comparative experiments and ablation experiments have demonstrated the effectiveness and generalization of the proposed method. CLRD-IIDM can adapt well to complex and ever-changing real-world network environments, and has certain practical value.

    参考文献
    相似文献
    引证文献
引用本文

金志刚,刘凯,陈旭阳,武晓栋.基于对比学习和关系蒸馏的增量式入侵检测方法[J].哈尔滨工业大学学报,2026,58(9):60. DOI:10.11918/202407062

复制
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-07-22
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2026-10-11
  • 出版日期:
文章二维码