Machine Learning for video-conference traffic classification

In this day and age, more and more people are using Online meeting tools to do meetings from the comfort of their office. However, the quality of these meetings as perceived by the users still requires some work, in order to make the experience comparable with the live one. So, policies for network management that favour Quality of Experience (QoE) of users are needed ever-more. The thesis aims to move the first step towards QoE defining for the first time the concept of user experience applied to this field, exploiting the potential of the assets based on machine learning. To do this, we explore the Real-time traffic (RTP) that is generated by meeting software and classify the different traffic flows – audio, video etc.- quality, through several classification algorithms. The thesis is organized as follows: first, we focus on Data collection which is useful to create a large dataset of diverse online meeting software traffic. Then we face the Feature extraction, where we calculate statistics on the packets, per RTP flow, per second. In the end, features are used as input to our multiclass classification algorithm and the results are analysed. The goal is to classify the flows into 6 classes: Audio, FEC audio, LQ video, MQ video, HQ video, FEC video and Screen sharing. 1. Data collection At the beginning we use the meeting tool Webex Meetings to collect data, performing active experiments with different settings: varying the quality of video and the network conditions. Using web scraping and web programming, we develop a tool that can automate the process of data collection, simulating human behaviour in the calling action. In this way, we are able to collect around 40 hours of various types of calls embracing all the classes. 2. Feature extraction After the first phase, it is time to synthesize the collected packets aggregating the various RTP flows per seconds and calculating statistics on different features such as length of the packets, bit rate, interarrival time, RTP timestamps etc. We engineer these features to recognize the different flows, so some of them are personalized for a certain class, for example, FEC audio or FEC video. iv 3. Multi-class Classification In the end, we train, test and compare three different algorithms to obtain the perfect match, based on a long grid search process aiming to find the best tuning parameters. The choice of algorithms is based on the different qualities, first trying a simpler method such as the K-NN based on a distance metric, then more complex methods such as random forest and feed-forward neural network using an entropy-based metric. The results obtained are very good, all with an accuracy of around 0.9. Random Forest turns out to be the best to classify the traffic with an accuracy of 0.93, follows k-NN with 0.89 and finally the Neural network with a value of 0.87 because of its complicated settings. Also an easy algorithm like K-NN in this case give good results, this is due to the fact that the data collection done and features construction have been executed in a precise way.