深度学习驱动的SDN网络应用智能分类

192 浏览量更新于2024-07-15 1 收藏 973KB PDF 举报

本文档探讨的主题是"基于深度学习的SDN网络应用分类"，发表于2017年11月5日，并在2018年2月进行了修订，最终接受于2月21日。该研究论文的DOI为10.1002/ett.3302，显示出其在学术界的专业性和权威性。SDN（Software-Defined Networking，软件定义网络）作为一种新兴的网络架构，通过将控制平面与数据平面分离，实现了网络功能的灵活编程。研究的主要内容聚焦于利用深度学习技术对SDN网络中的各类应用进行自动分类。深度学习是一种人工智能技术，特别适用于处理大量复杂数据，如网络流量、协议特征等。在SDN环境中，这些数据可以用于识别和区分不同的应用程序，例如视频流、语音通信、文件传输等，这对于网络优化、性能监控和安全策略制定具有重要意义。论文的作者包括来自东北大学计算机科学与工程学院的郑创创、王兴伟、李富良、何强以及信息科学与工程学院的黄敏教授，分别来自软件学院和过程工业综合自动化国家重点实验室。这表明研究团队具有深厚的技术背景和跨学科合作的优势。论文可能包含以下关键知识点： 1. **深度学习基础**：介绍深度神经网络（DNNs）的基本原理、结构和训练方法，以及它们在处理网络流量数据方面的优势。 2. **SDN网络架构**：简述SDN的关键组件，如OpenFlow等，以及如何通过控制器实现网络功能的虚拟化。 3. **网络应用特征提取**：阐述如何从网络流量中提取有意义的特征，以便深度学习模型能够理解和区分不同的应用类型。 4. **分类模型设计**：描述构建的深度学习模型，可能是卷积神经网络（CNN）、循环神经网络（RNN）或其他适合网络数据的模型。 5. **实验与评估**：详述实验设置，包括数据集的选择、模型训练过程、性能指标（如准确率、召回率等）以及结果分析。 6. **应用实例与讨论**：展示在实际SDN网络中的应用案例，探讨深度学习分类方法的实际效果和可能的应用场景。 7. **未来研究方向**：提出基于深度学习的SDN网络应用分类领域的潜在研究挑战和未来发展展望。这篇研究论文为深入理解SDN网络的动态行为，利用深度学习技术提升网络应用分类的精度和效率提供了有价值的理论和实践支持。它对于推动SDN技术在智能网络管理和优化中的应用具有重要价值。

4of18 ZHANG ET AL.

flow feature selection architecture was composed of flow feature manager model and flow feature selector model. The

former was used to obtain additional flow features, whereas the latter was responsible to analyze and select flow features.

As described in the introduction, the distinguishing features of SDN are helpful to achieve the application classification.

Some researchers in other works

9-11

have considered utilizing the SDN architecture for network application classification.

Prasad and Kataoka

designed an application-aware multipath packet forwarding mechanism by combining machine

learning and SDN. To achieve the application awareness, the C4.5 decision tree algorithm was used to build the application

classifier, and the machine learning–based trainer and classifier were integrated in the controller because of a global view

and the logically centralized control capability. Amaral et al

presented a novel machine learning–based data collection

and traffic classification architecture, which could be applied to legacy network and SDN network. In this architecture,

the controller collected the flow statistics from the switches and then used the s upervised machine learning algorithm to

classify traffic. Different from most research works on application classification, to achieve the classification and s atisfy

the QoS requirement of the network application at the same time, Wang et al

devised a QoS-aware traffic classification

framework in SDN. In this framework, deep packet inspection technology was used to detect the elephant flow, and the

semisupervised machine learning algorithm was used to achieve the QoS-aware traffic classification through the mapping

function. Specifically, the application flow was mapped to a certain predefined QoS class according to flow features in the

mapping function.

Although the aforementioned research works take the SDN architecture into account for application classification,

these classification methods merely belong to traditional shallow learning network. It cannot efficiently deal with mas -

sive data because of the limited feature learning ability. Compared with shallow learning network, deep learning network

has more powerful feature learning ability and can achieve deeper features from data through the training.

Deep

neural network has been widely used for intrusion detection, traffic flow prediction, and application classification.

26-30

Salama et al

proposed a hybrid intelligent anomaly intrusion detection scheme by jointly using the RBM-based deep

belief network and SVM. In this intrusion detection scheme, the deep belief network was used to reduce the feature

dimensionality, and SVM was used to classify whether the network traffic is normal. Similar with this work, Fiore et al

directly applied the RBM classifier to network anomaly detection by using the s emisupervised learning method. To deal

with massive network traffic, Lv et al

proposed a novel traffic flow prediction method on the basis of deep learning. In

this method, the stacked autoencoder model was used to obtain flow features since it could extract inherent features on

the basis of the data set from the lowest level to the highest level. To improve the forecasting accuracy of network traffic

and solve the congestion problem, Yang et al presented a stacked autoencoder Levenberg-Marquardt model.

The pro-

posed deep network was used to design an optimized structure and obtain flow features by means of the greedy layerwise

training algorithm. Huang et al

designed a novel deep network architecture for traffic flow prediction in transportation

research. The deep neural network was composed of a DBN and a multitask regression layer. The former was designed to

learn effective features with the greedy layerwise training algorithm, whereas the latter was used to predict the traffic flow.

Inspired by the aforementioned research works, in this paper, we propose an application classification framework by

combing SDN and deep learning. With the powerful computing capability, we use the controller to deal with the massive

network traffic and flow statistics. We construct a hybrid deep learning network model, which is composed of the stacked

autoencoder and the softmax regression layer, to extract flow features and build the application classifier.

3 FRAMEWORK DESIGN

In this section, we propose a novel framework to collect network traffic, process flow statistics, and classify network

application by using SDN and deep learning technology. Its main goal is to collect and process flow statistics information

and build an application classifier that divides different types of application flows into different categories.

There are several definitions on the flow. In this paper, we use the socket to define a flow, by a quintuple, that is, ⟨source

IP address, destination IP address, transport protocol (eg, Transmission Control Protocol (TCP) and User Datagram

Protocol ), source port number, destination port number ⟩.

Our proposed application classification framework is shown in Figure 1. As depicted in Figure 1, the control plane con-

sists of four important function modules (ie, network monitoring module, flow statistic module, data processing module,

and deep learning classifier module) and a database, which stores flow features information. The 4 function modules are

described as follows in details.

Network monitoring module: It is responsible to gather and analyze network traffic information from underlying

switches to obtain some useful flow information (eg, protocol type, source port, and destination port number of the flow).

剩余17页未读，继续阅读

weixin_38656142

粉丝: 6
资源: 909

深度学习驱动的SDN网络应用智能分类

SDN环境下不同机器学习算法的网络流量分类分析.pdf

深度解析SDN

基于深度学习的SDN恶意应用的检测方法.pdf

基于切片的深度学习SDN恶意应用程序的检测方法.pdf

基于深度学习的软件定义网络应用策略冲突检测方法

基于深度学习的软件定义网络应用策略冲突检测方法.pdf

SDN下基于深度学习混合模型的DDoS攻击检测与防御.pdf

一种基于深度强化学习的SDN路由算法.pdf

基于图卷积神经网络的SDN网络流量预测.pdf

基于深度学习的流量工程算法研究与应用.docx

最新资源