Summary: | 碩士 === 國立臺北科技大學 === 資訊工程系 === 107 === As networks grow increasingly complex, the flexible configuration of SDN and its application received more and more approved. For example, the 3GPP Release 15 NSA 5G NR specification selected SDN as its backbone architecture and developed network slicing technique to adapt to diverse requirements in terms of functions and services. In addition, the emerging Internet-of-Thing (IoT) is expected to bring huge amount of data over networks, which may introduce traffic congestion and impact to user experience.
This research associated SDN with Deep learning and designed and realized a system: packet/flow identification. Its objective was to enabling QoS SDN transport. To the design methodology of systems, paper applied semi-supervised learning–Convolutional Neural Network (CNN), respectively. The system first adopted supervised machine learning with open data to got the packet identification model /algorithm and set up the corresponding rules to an external device. Then, the identification process could be done without involution of controller. Doing this way, the load of controller could be mitigated. Different from many other approaches using open data only, to adapt realistic network operation, this proposal collected unrecognized packets thru switches/controller and sent back to DLS (Deep Learning server) to update the identification model. With this semi-supervised machine learning operated constantly the system could dynamically catch up new type of packet and keep the neural networks model up-to-date. Consequentially, this continuing iteration would raise the accuracy of packet identification.
|