Network traffic classification with machine learning and edge computing

随着增强现实、无线多媒体、云计算和物联网等数据密集型应用的日益流行，当前一代网络基础设施将面临严峻的挑战，以满足爆炸性的服务需求。为了在这些不同的应用之间有效地分配可用的网络资源，流量分类已经成为网络优化运行的关键技术。然而，传统的分类方法，如基于网络协议接口或报文内容检测的分类方法，已不再适合满足支持现代应用和服务的准确性、时延和隐私要求。新应用的不断变化的特点和不断增长的通信量要求一种灵活和自动的方法。在这个项目中，我们将应用动态机器学习技术来识别和分类互联网服务提供商的网络流量。该项目的一个独特特点是，我们将利用来自TELUS的大量匿名用户流量数据，调查监督和非监督学习的混合组合。我们还将利用网络边缘新兴的计算能力，即网络流量更本地化，在本地用户之间共享共性，以提高流量分类的准确性。通过数学分析、计算机模拟和大规模数据实验，我们将生成关于如何设计和操作网络流量分类器的实用指南，以最佳地平衡准确性、保密性、成本和延迟之间的权衡。通过促进工程理论、技术方法和标准化政策，促进系统改进、成本降低、可持续增长和长期竞争力，拟议的研究结果预计将使我们的行业合作伙伴和整个加拿大信息和通信业受益。

With the emerging prevalence of data-hungry applications, such as augmented reality, wireless multimedia, cloud computing, and Internet of Things, the current generation of network infrastructure will face severe challenges in its struggle to satisfy the exploding service demands. In order to efficiently allocate the available networking resources among these diverse set of applications, traffic classification has become an essential technology for the optimal operation of a network. However, classical classification methods, such as those based on network protocol interface or packet content inspection, are no longer suitable to meet the accuracy, delay, and privacy requirements to support modern applications and services. The ever changing characteristics of new applications and increasing volume of traffic demand a flexible and automatic approach. In this project, we will apply dynamic machine learning techniques to identify and categorize the network traffic of an Internet service provider. A unique feature of this project is that we will leverage the availability of a vast amount anonymous user traffic data from TELUS, to investigate into a hybrid combination of both supervised and unsupervised learning. We will also take advantage of the emerging capabilities of computing at the network edge, where the network traffic is more localized with shared commonalities among local users, to improve the accuracy of traffic classification. Through mathematical analysis, computer simulation, and large-scale data experimentation, we will generate practical guidelines on how to design and operate network traffic classifiers, to optimally balance the tradeoff between accuracy, privacy, cost, and delay. The outcomes of the proposed research are expected to benefit both our industry partner and the Canadian information and communication industry at large, by promoting engineering theory, technical methods, and standardization policies that can lead to system improvement, cost reduction, sustainable growth, and long-term competitiveness.