為符合使用者越來越高的網路服務品質，網際網路從傳統基本的路由機制發展至今，逐漸演變為服務導向（service-oriented）的架構。為使網路更加彈性，許多新興的網路虛擬化技術，如網路功能虛擬化（Network Function Virtualization，NFV）與軟體定義網路（Software Defined Network，SDN）因此應運而生。而在服務導向的架構下，快速地、有效地在網路中分類封包便是一項重要的技術。藉由SDN中的控制器（controller）可程式化、中樞管理的優勢，網路管理者可在controller中撰寫程式，部署封包分類機制與路由決策。而以SDN的技術為基礎，廣域軟體定義網路（Software Defined Wide Area Network，SD-WAN）近年逐漸成為主流架構。在SD-WAN中，多種服務有多條路由可以選擇，且不同的服務對其網路品質的要求也有所不同，因此服務的流量分類（traffic classification）與路由決策（routing decision）是SD-WAN中不可或缺的部分。本論文為了在SD-WAN架構中達到流量分類與路由決策，提出兩個主要機制：CAPC，基於近年發展快速的卷積神經網路（Convolutional Neural Network，CNN）與自編碼（Deep Autoencoder，DAE）技術有效分類自行錄製與公開加密資料集的服務，分別達到優於其他深度學習模型的99.98%與97.42%的最高正確率； RED-STAR，基於深度強化學習（Deep Reinforcement Learning，DRL）技術，將CAPC所分類出的服務類型與機制中所正規化的網路指標作為DRL中神經網路的輸入值，逐漸學習並定期計算出每個服務應被導向何路由才能取得最大獎勵值（reward），並在實際服務流量情境下得到高於其他路徑分配演算法的1.7327和1.8642平均reward。 ;To meet users’ requirements for higher service quality, the Internet has been evolving from the traditional routing mechanism to the current service-oriented architecture. To make networks more flexible, numerous emerging network virtualization technology, such as Network Function Virtualization (NFV) and Software Defined Network (SDN) were proposed. In the service-oriented architecture, instantly and effectively classifying packets during transmission is essential. With the advantages of programmability and centralized control of SDN controller, network managers can program within the controller, and deploy packet classification and routing mechanism. Based on SDN, Software Defined Wide Area Network (SD-WAN) has become the main stream of network architecture in recent years. In SD-WAN, there are plenty of network services having to be routed to several paths, and different services might even have different service quality demands. As a result, traffic classification and routing decision are indispensable in SD-WAN. To classify traffic and route service traffic, two mechanisms have been proposed: CAPC, which classifies packets of different kinds of services based on the recent fast-growing Convolutional Neural Network (CNN) and Deep Autoencoder (DAE), with the highest accuracies of 99.98% and 97.42% on the self-captured traffic and the open encrypted dataset; RED-STAR, based on Deep Reinforcement Learning, takes the classification results from CAPC and the regularized network metrics as input values for DRL, and learns to decide the path to which a service should be routed for the maximum reward value, with the average reward values of 1.7327 and 1.8642 in the real-case scenarios.
