| 基于掩码自监督学习的CONV-BiLSTM入侵检测方法 |
 点此下载全文 |
| 引用本文：吴世钜,柳毅.基于掩码自监督学习的CONV-BiLSTM入侵检测方法[J].计算技术与自动化,2026,(1):164-170 |
| 摘要点击次数: 227 |
| 全文下载次数: 187 |
|
|
| 中文摘要:针对入侵检测数据集分类数量不平衡问题，提出了一种基于掩码自监督学习的CONV-LSTM入侵检测方法。首先对数据训练集部分标签进行屏蔽，提高模型预测性能。然后通过卷积注意力自编码对数据特征进行降维和重构，提取数据空间特征；再结合双层Bi-LSTM提取数据时间特征，实现由粗粒度到细粒度的学习。同时针对数据集部分攻击种类数量过少的情况，提出使用均衡化损失v2损失函数进行训练，来增加少数种类权重，提高其预测召回率。通过对UNSW-NB15数据集的实验结果表明，该方法在二分类和多分类任务准确率达到96.57%和86.46%。对比其他方法，本文提出的方法在准确率、检测率、误报率以及数据各个分类的召回率表现更好。 |
| 中文关键词:入侵检测 卷积自编码器 注意力机制 掩码自监督学习 |
| |
| CONV-BiLSTM Intrusion Detection Method Based on Mask Self-supervised Learning |
|
|
| Abstract:A CONV-LSTM intrusion detection method based on mask self-supervised learning is proposed to address the issue of imbalanced classification in intrusion detection datasets. Firstly, some labels in the training set are masked to improve the predictive performance of the model. Then, convolutional attention autoencoder is used to reduce dimensionality and reconstruct data features, extracting spatial features from the data. Combined with dual layer Bi-LSTM to extract temporal features of data, achieving learning from coarse to fine granularity. At the same time, in response to the situation where the number of attack types in the dataset is too small, it is proposed to use the balanced loss v2 loss function for training to increase the weight of a few types and improve their prediction recall rate. The experimental results on the UNSW-NB15 dataset show that the accuracy of this method in binary and multi classification tasks reaches 96.57% and 86.46%, respectively. Compared to other methods, the method proposed in this article performs better in accuracy, detection rate, false positive rate, and recall rate for various categories of data. |
| keywords:intrusion detection convolutional autoencoder attention mechanism mask self-supervised learning |
| 查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|
位访问者,今日访问
