基于深度置信网络的入侵检测研究

Research on Intrusion Detection Based on Deep Confidence Network

摘要: 入侵检测作为一种积极主动的安全防范技术，在网络安全上的应用历来已久。但是随着互联网的发展和应用的不断深化，网络攻击和入侵在数量和技术水平上的不断变化，基于新类型、多并发的攻击，使传统的入侵检测技术已经无法满足现有网络安全的要求。深度学习作为目前机器学习和人工智能的前沿技术，在语音识别、计算机视觉、大数据处理等方面都取得了巨大成果，也为解决当前的入侵检测问题提供了一个新的思路。本文基于对传统入侵检测技术的研究，结合深度学习方法下的深度置信网络，提出了一种基于深度置信网络的入侵检测技术，根据入侵检测数据的特点对数据进行过取样和非[0, 1]区间的归一化，在深度置信网络的参数更新过程中，采用批梯度下降的可变学习率算法，加快了参数的更新过程，并在每批训练数据中，加入了对少类别标签的区分度，提高了准确率。实验证明，利用本文提出的方法，可以很好地提高入侵检测的准确率。

Abstract: As an active security prevention technology, intrusion detection has been used in network security for a long time. But with the development and application of Internet, network attack and intrusion are constantly changing in quantity and technology level. Based on new types and concurrent at-tacks, traditional intrusion detection technology has been unable to meet the requirements of existing network security. As a frontier technology of machine learning and artificial intelligence, deep learning has made great achievements in speech recognition, computer vision and big data processing, and has also provided a new idea for solving the current intrusion detection problem. This paper studies the traditional intrusion detection technology based on learning method combined with the depth and the depth of the belief network, proposes an intrusion detection technology based on deep belief networks, according to the characteristics of intrusion detection data of over sampling and non [0, 1] interval of data normalization, updates the parameters in the deep belief network in the process of using variable number of the gradient descent algorithm to speed up the learning rate, the parameters of the update process, and in each batch of the training data, join the discrimination on the labels, and improve the accuracy. Experiments show that the accuracy of intrusion detection can be improved greatly by using the method proposed in this paper.