Abstract: In view of the great achievements of convolutional neural networks in many fields such as computer vision, a method of applying multi-scale convolutional neural networks to the field of network intrusion detection is proposed. This method converts the network data in IDS into data that the convolutional neural network can input, uses different scales of convolution to verify a large number of high-dimensional unlabeled original data for different levels of feature extraction, and then uses the BN method to optimize the learning rate of the network structure. The optimal feature representation of raw data. Experiments using the KDDcup99 data set for experimental testing, compared with the classic model, the results show that the MSCNN model not only has a fast convergence rate, but also the false detection rate is reduced by 4.02% on average, and the accuracy rate is increased by 4.38% on average. Therefore, the MSCNN method is a feasible and efficient method and provides a brand-new idea for the field of network intrusion detection systems.

Key words: intrusion detection, deep learning, convolutional neural networks, BN algorithm, multiscale convolution

摘要： 鉴于卷积神经网络在计算机视觉等诸多领域取得的巨大成就，提出一种将多尺度卷积神经网络应用到网络入侵检测领域的方法。该方法将IDS中的网络数据转化成卷积神经网络能够输入的数据，利用不同尺度卷积核对大量高维无标签原始数据进行不同层次特征提取，再采用BN方法优化网络结构学习率，从而获得原始数据的最优特征表示。实验采用 KDDcup99数据集进行实验测试，与经典的模型相比，结果表明MSCNN模型不仅收敛速度快，而且误检率平均降低4.02%，准确率平均提高4.38%。因此MSCNN方法是一种可行且高效的方法，为网络入侵检测系统领域提供一种全新的思路。

关键词: 入侵检测, 深度学习, 卷积神经网络, BN算法, 多尺度卷积