Real-time communication (RTC) platforms have seen a considerable surge in popularity in recent years, largely due to the COVID-19 pandemic which facilitated remote work. To ensure adequate Quality of Experience (QoE) for users, a good congestion control algorithm is needed. RTC applications use UDP, so congestion control is done on the application layer, leaving way for advanced algorithms. In this paper, we propose ReCoCo, a solution for congestion control in RTC applications based on Reinforcement learning (RL). ReCoCo gains information about the network conditions at the receiver-side, such as receiving rate, one-way delay and loss ratio and predicts the available bandwidth in the next time bin. We train ReCoCo on 9 bandwidth trace files that cover a vast array of network types. We try different algorithms, states and parameters, training both specific and general models. We find that ReCoCo outperforms the de-facto standard heuristic algorithm GCC in both specialized and general models. We also make observations on the difficulty of generalization when using RL.

Real-time communication (RTC) platforms have undergone a consistent increase in popularity in recent years, and nowadays, they are fundamental for both work and leisure purposes. To ensure adequate Quality of Experience (QoE) for users of RTC services, we need proper traffic management policies, that, when critical network conditions are detected, react by operating either at the network configuration level or on the application to improve QoE. However, predicting critical network conditions, especially packet losses that are particularly harmful to QoE, is a very challenging task. In this paper, we propose a system for predicting packet losses that might occur in the near future (i.e., in a second) for RTP streaming traffic. We analyze several ML algorithms, from standard techniques to deep neural networks and anomaly detection algorithms, and we apply them to more than 66 hours of data from two popular RTC applications. The selection of the algorithm and its tuning turn out to be fundamental to achieving good performance. In one of the best settings, which are based on a Balanced Random Forest classifier, we obtain a recall of 0.82.

Real-time communication (RTC) applications have become largely popular in the last decade with the spread of broadband and mobile Internet access. Nowadays, these platforms are a fundamental means for connecting people and supporting businesses that increasingly rely on forms of remote work. In this context, it is of paramount importance to operate at the network level to ensure adequate Quality of Experience (QoE) for users, and appropriate traffic management policies are essential to prioritize RTC traffic. This in turn requires the network to be able to identify RTC streams and the type of content they carry. In this paper, we propose a machine learning-based application to classify media streams generated by RTC applications encapsulated in Secure Real-Time Protocol (SRTP) flows in real-time. Using carefully tuned features extracted from packet characteristics, we train models to classify streams into a variety of classes, including media type (audio/video), video quality, and redundant streams. We validate our approach using traffic from over 62 hours of multi-party meetings conducted using two popular RTC applications, namely Cisco Webex Teams and Jitsi Meet. We achieve an overall accuracy of 96% for Webex and 95% for Jitsi, using a lightweight decision tree model that makes decisions based solely on 1 second of real-time traffic. Our results show that models trained for a particular meeting software have difficulty when used with another one, although domain adaptation techniques facilitate the transfer of pre-trained models.

Real-time communication (RTC) platforms have become increasingly popular in the last decade, together with the spread of broadband Internet access. They are nowadays a fundamental means for connecting people and supporting the economy, which relies more and more on forms of remote working. In this context, it is particularly important to act at the network level to ensure adequate Quality of Experience (QoE) to users, where proper traffic management policies are essential to prioritize RTC traffic. This, in turn, requires in-network devices to identify RTC streams and the type of content they carry. In this paper, we propose a machine learning-based application to classify, in real-time, the media streams generated by RTC applications encapsulated in Secure Real Time Protocol (SRTP) flows. Using carefully tuned features extracted from packet characteristics, we train a model to classify streams into an ample set of classes, including media type (audio/video), video quality and redundant streams. To validate our approach, we use traffic from more than 88 hours of multi-party meeting calls made using the Cisco Webex Teams application. We reach an overall accuracy of 97% with a light-weight decision tree model, which makes decisions using only 1 second of traffic.

Real-Time Communication (RTC) applications have become ubiquitous and are nowadays fundamental for people to communicate with friends and relatives, as well as for enterprises to allow remote working and save travel costs. Countless competing platforms differ in the ease of use, features they implement, supported user equipment and targeted audience (consumer of business). However, there is no standard protocol or interoperability mechanism. This picture complicates the traffic management, making it hard to isolate RTC traffic for prioritization or obstruction. Moreover, undocumented operation could result in the traffic being blocked at firewalls or middleboxes. In this paper, we analyze 13 popular RTC applications, from widespread consumer apps, like Skype and Whatsapp, to business platforms dedicated to enterprises - Microsoft Teams and Webex Teams. We collect packet traces under different conditions and illustrate similarities and differences in their use of the network. We find that most applications employ the well-known RTP protocol, but we observe a few cases of different (and even undocumented) approaches. The majority of applications allow peer-to-peer communication during calls with only two participants. Six of them send redundant data for Forward Error Correction or encode the user video at different bitrates. In addition, we notice that many of them are easy to identify by looking at the destination servers or the domain names resolved via DNS. The packet traces we collected, along with the metadata we extract, are made available to the community.