卷积神经网络图像识别技术在入侵监测系统中的应用研究  

Application Study of Convolution Neural Network Image Recognition Technology in Intrusion Detection System

作　　者：王洪德[1,2] 王焕青 WANG Hongde WANG Huanqing(School of Civil ＆ Safety Engineering, Dalian Jiaotong University, Dalian 116028, China Tunnel ＆ Underground Structure Engineering Center, Dalian Jiaotong University, Dalian 116028, China)

机构地区：[1]大连交通大学土木与安全工程学院,辽宁大连116028 [2]大连交通大学隧道与地下结构工程技术研究中心,辽宁大连116028

出　　处：《大连交通大学学报》2017年第5期107-110,共4页Journal of Dalian Jiaotong University

基　　金：国家自然科学基金资助项目(U1261121/E0422);中国铁路总公司科技研究开发计划资助项目(2015Z002)

摘　　要：为应对入侵手段复杂多样的安全形势,解决现有入侵监测技术成本高、适应性差的问题,设计研发一种成本低廉、兼容性好、方便拓展的入侵监测系统.应用开源深度机器学习框架Tensor FLow实现基于卷积神经网络的图像识别算法,构建辅助安全系统,并在多种工况下对系统有效性进行验证.结果表明:系统能以较高准确率对入侵行为进行识别,在多分类情景下,随训练样本数量的增加,模型预测准确率得到提高,而收敛时间有所增加,开启GPU加速后缩短为原来的1/10.In order to solve the complicated and diversified security situation of intrusion and to solve the prob- lem of high cost and poor adaptability of existing intrusion detection technology, an intrusion detection system with low cost, good compatibility and convenient expansion is designed and developed. The image recognition algorithm is realized based on convolution neural network using open-source depth machine learning framework TensorFLow, and the system architecture is constructed. Then the effectiveness of the system is validated under various operating conditions. Results show that the intrusion behavior can be identified by the system with high accuracy. In a multi-class scenario, with the increase of the number of training samples, the prediction accuracy of the model is increased. While the convergence time is increased, the time is shorten to one tenth when the GPU acceleration is turned.

关 键 词：入侵监测 图像识别 卷积神经网络 TENSOR FLOW 

分 类 号：TN948.6[电子电信—信号与信息处理] TP391.41[电子电信—信息与通信工程]