Multimedia traffic classification based on discrete time markov chains

Official URL: https://risc01.sabanciuniv.edu/record=b3147292_(Table of contents)

Traffic prioritization has recently become critical for home Wi-Fi networks due to the increased number of connected devices and wide variety of applications. While some of these applications are delay sensitive, some have high throughput requirements. Therefore, managing different traffic types adaptively with regard to their requirements is crucial for a better Quality of Experience (QoE). Traffic type classification can be used to that end for detecting the specific requirements and enhance the Quality of Services (QoS). In the scope of the thesis, we propose to model the multimedia traffic flow as a stochastic discrete-time Markov chain (DTMC) in order to take into account the strong sequentiality (i.e. the dependencies across the data instances) in the traffic flow observations. Within that approach four novel data-driven classification schemes are presented. The first one is k Nearest Markov Components (kNMC) which relies on a Markov modeling of bit-rate. kNMC considers a mixture of Markov components and classifies by using a log-likelihood-based distance between train and test instances. The second classifier is kNMC-3D which applies the same technique with kNMC but focuses on the number of packets and inter-arrival times besides bit-rate. The other two classifiers exploit Error Correcting Output Codes (ECOC) for solving the multiclass problem with multiple binary kNMC-3D classifiers. The third classification scheme namely Confusion Based ECOC (CB-ECOC) proposes a custom-designed ECOC matrix which addresses the errors of kNMC-3D. The fourth classifier named as 2-Level ECOC, adds another classification level to CB-ECOC iv for resolving the Skype identification issue of CB-ECOC. Considering multimedia data from popular applications such as Youtube, Netflix, Skype, Whatsapp, and Spotify from our introduced dataset, average traffic type classification accuracies are obtained up to 96.15% at the application level. Considering the given applications in different traffic categories, such as, Video on Demand (VOD), Sharing and Media Screening (S&MS), Video Live Streaming (VLS), and Teleconferencing (TC), average classification accuracies up to 97.75% are reached at the category level. The presented classifiers are also evaluated with the benchmark dataset from the literature and average classification accuracies are observed up to 97.75% at the application level, and up to 99.59% at the category level. In our extensive experiments, we observed that the introduced classifiers are highly accurate as compared to prominent competitors such as Support Vector Machines (SVM), Random Forest (RF), autoencoders and problem-independent ECOC models, e.g, One Versus One (OVO) and One Versus All (OVA).