| 摘要: | 隨著 5G 時代的到來，其中兩項關鍵技術為軟體定網路以及機器學習也越發成熟：
透過軟體定義網路，網路業者能更有效且彈性的進行網路功能的分配以及拓樸彈性調
整。另一方面，透過機器學習對於不同的學習方式以及學習任務的框架分類，讓機器
學習能應用在不同許多領域，以相對人為更短的時間，去對資料進行分析並部署對應
策略。但是也同時因為 5G 大數據時代，兩領域也相繼產生出新問題及挑戰：軟體定義
網路方面，以最常見的 OpenFlow 協定為例，為了適應不同的風包格式，標頭已經越
來越複雜，部署也愈發複雜；機器學習方面，因為大量數據的學習，收斂時間開始被
拉長，且關於隱私數據問題也一直無法被有效解決，故在模型表現方面也需要進一步
的去改善。
藉由 P4 交換機的發明，把原本從下而上的交換機設計，改成了由上而下，其好處
是交換機的協議不再跟硬體做綁定，進而對封包做更有效以及彈性的動作規則更改。
在近幾年興起的聯邦式學習框架，不只可以有效保留終端隱私且仍能進行有效的模型
訓練，克服了以前集中式以及分散式學習架構上的問題。本研究也將利用上述 P4 和聯
邦式學習的優點，設計一個進行物聯網封包傳輸的優化之環境研究，經由實驗結果可
以發現，與對照組三組環境相比，本論文所採用數據為最優，故能證明此兩項技術能
有效改善物聯網封包傳輸時之表現。
;With the advent of the 5G era, two key technologies, Software-Defined Networking
(SDN) and Machine Learning (ML), are becoming more and more mature. Through
SDN, network operators can allocate network functions and adjust topology flexibility
more effectively and flexibly. On the other hand, by classifying different learning styles
and learning tasks by ML, it can be applied in many different fields, and the data can be
analyzed and corresponding strategies deployed in a relatively short time. However, at
the same time, because of the 5G big data era, new problems and challenges have emerged
in the two fields. In terms of SDN, taking the most common protocol OpenFlow as an
example, in order to adapt different packets formats, the header has become more and
more complex, and the deployment has become more and more complex; In the aspect of
ML, the convergence time begins to be lengthened because of the learning large amount
of data, and the problem of private data has not been effectively solved, so the model
performance needs further improvement.
With the invention of P4 switch, the original Bottom-Up switch design has been
changed to Top-Down, which has the advantage that the protocol of the switch is no
longer bound to the hardware, thus making more effective and flexible changes to the
action rules of packets. In recent years, the emerging Federated Learning (FL) framework
can not only keep the privacy of the terminal effectively, but also carry out effective model
training, overcoming the problems of the previous centralized and decentralized learning
framework. This study will also make use of the advantages of P4 and FL mentioned
above to design an environment to optimize the packet transmission of the Internet of
Things (IoT). Through the experimental results, it can be found that compared with the
control group, the environment set in this paper is the best, so it can be proved that these
two technologies can effectively improve the performance of packet transmission of the
IoT. |
