Abordagem Híbrida FuzzyNetClass: Uma Contribuição à Classificação do Tráfego de Streaming de Vídeo Integrando Lógica Fuzzy Valorada Intervalarmente e Aprendizagem de Máquina

Abstract

mong the challenges for classifying network traffic, the maintenance of user pri vacy stands out based on cryptography and traffic obfuscation techniques. This chal lenge, associated with the dynamics and similarity between the protocols used by the different services, as well as the typical oscillations in the operational conditions of the networks, which impact the analysis of the captured data, has made the use of clas sical methods for classification unfeasible. Considering this, the general goal of this Thesis is the conception of a hybrid approach for network traffic classification, called FuzzyNetClass, directed to the current profile of use of computer networks and thus considering the uncertainties generated by fluctuations in the network resources of the infrastructures shared, which are non-deterministic. More specifically, this work aims to contribute to the traffic classification related to video streaming, exploring the inte gration of inference systems based on interval-valued fuzzy logic and machine learning algorithms. In this perspective, the FuzzyNetClass approach extends the related work by exploring learning algorithms that contribute to the classification of the proposed fuzzy control system, preserving aspects related to its interpretability for the involved specialists. Among the FuzzyNetClass approach contributions, the following stand out: (i) classification based on the specialist’s knowledge and the exploration of multivalued fuzzy logic; (ii) hybrid classification, which has the potential to provide more reliable and more realistic results; and, (iii) the design of optimization mechanisms that come from the integration between machine learning techniques and inference based on multivalued logic. Three case studies were discussed to evaluate the FuzzyNetClass approach’s contributions, considering datasets conceived from real network traffic cap tures. The results were promising and pointed to continuing study and research efforts.