Descobrindo perfis de tráfego de usuários: uma abordagem não supervisionada

• Ano de defesa: 2019
• Autor(a) principal: Streit, Ananda Görck
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia de Sistemas e Computação; UFRJ
• Palavras-chave em Português: Redes de acesso residencial; Perfis de tráfego; Aprendizado de máquina; CNPQ::ENGENHARIAS
• Link de acesso: http://hdl.handle.net/11422/13552

The increasing complexity of home networks calls for novel strategies towards e cient network management and workload characterization. In particular, understanding the characteristics of the tra c generated by users is of paramount importance for network planning. Previous work focuses primarily on Deep Packet Inspection (DPI) and/or considers pre-determined patterns to classify tra c ows and detect the application being accessed by users. In this work we use unsupervised machine learning techniques with the objective of discovering users’ tra c pro les. In partnership with an Internet Service Provider (ISP) we collected the download and upload tra c of more than 2,000 home routers of the ISP clients. We then use a tensor decomposition technique (PARAFAC) to extract relevant features from our network traces. We show how the results of PARAFAC and a hierarchical clustering algorithm simplify the task of grouping time series with similar daily tra c patterns. We also show how new users can be classi ed from the decision tree obtained with clustering. To characterize users’ behavior over periods longer than a day, we use the information of the clusters and a Hidden Markov Model (HMM). The results indicate that users tend to maintain a speci c pattern over time, facilitating network planning and management tasks.