Abstract:

In order to solve the problem that the traditional intrusion detection methods are difficult to quickly and accurately extract feature information from massive unlabeled network data to identify abnormal intrusions, this paper proposes an intrusion detection method of Softmax Classification based on the Improved Deep Belief Network（IDBN-SC）, where the deep belief network is employed for conducting unsupervised feature learning on the original network data and the adaptive learning rate is used to reduce the time required to train the network model. Moreover, the softmax classifier is utilized to identify the types of network attack. Testing on the NSL-KDD dataset shows that the IDBN-SC method not only improves the recognition accuracy by an average of 3.02%, but also reduces the softmax classifier training time by 5.58 s.

Key words: restricted Boltzmann machine, intrusion detection, deep belief network, softmax classifier, adaptive learning rate, feature learning

摘要：

针对传统入侵检测方法很难快速准确地从海量无标签网络数据中提取特征信息以识别异常入侵，提出了基于改进的深度信念网络的softmax分类（IDBN-SC）入侵检测方法。利用改进的DBN对原始网络数据进行无监督特征学习，引入自适应学习速率减少训练网络模型所需要的时间；采用softmax分类器对获得的降维数据进行网络攻击类型识别。在NSL-KDD数据集上进行测试，相比其他入侵检测方法，实验结果表明IDBN-SC方法不仅识别准确率平均提高3.02%，而且其softmax分类器训练时间平均缩短5.58 s。

关键词: 受限玻尔兹曼机, 入侵检测, 深度信念网络, softmax分类器, 自适应学习速率, 特征学习