基于深度神经网络的入侵检测方法  

Intrusion detection method based on deep neural network

作　　者：钱铁云[1,2] 王毅 张明明 刘俊恺 

机构地区：[1]武汉大学计算机学院,湖北武汉430072 [2]武汉大学软件工程国家重点实验室,湖北武汉430072 [3]国网江苏省电力公司信息通信分公司,江苏南京210013 [4]北京汇通金财信息科技有限公司,北京100094

出　　处：《华中科技大学学报（自然科学版）》2018年第1期6-10,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61572376);高等学校学科创新引智计划资助项目(B07037)

摘　　要：为改善传统机器学习技术解决海量网络数据和复杂入侵模式对信息网络的入侵检测的不足,提出一种基于深度神经网络的入侵检测方法.采用神经元映射卷积神经网络（NPCNN）为网络结构,使用较少的连接和参数,具有易于训练和泛化能力强等优点.在训练过程中,使用Re LU激活器作为非线性激活函数,采用Adam算法进行模型学习,从而避免了传统深度网络须进行预训练的过程.在NSL-KDD数据集上的实验结果表明：提出的方法较基于传统机器学习的入侵检测方法具有良好的特征表征学习和分类能力,且随着数据量的增大,模型的分类精度有较大的提升.To deal with difficulties for traditional machine learning methods to deal with huge volume of data and complexity of intrusion models,a deep neuron network method was proposed.The neuron project convolution neuron network（NPCNN） was adopted as the infrastructure with less connections and parameters,which was easy to train and generalize.During the training process,the rectified linear unit layer was used as the activation function,and the Adam algorithm was used to train the model so as to save the time for pre-training.Extensive experiments were conducted on a publicly available dataset NSL-KDD.The experimental results demonstrate that the proposed approach has good characteristics to characterize learning and classification ability,and with the increase of data quatity,the classification accuracy of the model has been greatly improved.

关 键 词：入侵检测 深度学习 卷积神经网络 连接 参数 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]