Teaching

Contact: nikolas.wehner@uni-wuerzburg.de In recent years, the Natural Language Processing (NLP) domain has been transformed severely by novel learning mechanisms like attention and novel Deep Learning models like the general Transformer, BERT, and GPT-3. The attention mechanism aligns input sequences with output sequences in a way that the importance of each input element can be estimated for an output element. This limits its applicability not only to NLP, but it could also be a highly interesting concept for the networking domain in general. The novel models like the Transformer make use of attention mechanisms in order to be able to perform for example neural machine translation with high accuracy. In this work, the applicability of the attention mechanism and for example the Transformer model shall be evaluated for the use case of (Web/Video) QoE monitoring. For this work, the data will be provided. The main task is the modelling of the network traces and the optimization of the learning processes. Therefore, various approaches for the modelling and for the feature extraction shall be tested. (Also available as Bachelor's thesis or Praktikum)

Contact: nikolas.wehner@uni-wuerzburg.de The goal of this work is to estimate Web QoE KPIs like the Real-User-Monitoring Speed Index (RUMSI) or Page Load Time (PLT) directly from encrypted network traces using flow-based information only. This requires a reasonable feature extraction and engineering process as well as the use of machine learning techniques. Therefore, different feature sets and models shall be tested, e.g., using the response times of flows and modelling the data as time series. These approaches are however subject to change. The data for this work will be provided.

Contact: nikolas.wehner@uni-wuerzburg.de Abstract: Nowadays, basically all videos are streamed via HTTP Adaptive Streaming (HAS). An important concept in terms of HAS is the Quality of Experience (QoE), which describes how users perceive video streaming. QoE fairness extends the concept of QoE by comparing the QoE of mutliple clients in a shared network. An important component influencing the QoE and consequently the QoE fairness is the HAS adaptation logic, which is responsible for downloading the segments with the goal to obtain the best possible video quality without causing stalling, i.e., playback interruption due to buffer underrun. The goal of this work is to implement an adaptation logic, currently implemented in a simulation, into a real video player and evaluating the QoE and QoE fairness of the adaptation logic when confronted with low-latency situations in real scenarios. (Also available as Bachelor's thesis)

Contact/Kontakt: nikolas.wehner@informatik.uni-wuerzburg.de Auch als Praktikum verfügbar. Due to the encryption of network traffic and the introduction of DNS over HTTPS, the task of identifying websites based on the corresponding network traffic got difficult. Nonetheless, ISPs want to know which kind of traffic passes their networks in order to optimize the QoE of their customers. The goal of this work is to create a model which is able to reliably predict the called websites during a session. Therefore, reasonable features have to be extracted from the data and afterward they have to be ingested into selected Machine Learning algorithms.

Contact/Kontakt: nikolas.wehner@informatik.uni-wuerzburg.de Impairments in the Web QoE, e.g., low page load times, can lead to annoyed users which in the end can result in revenue losses for companies. In literature, several algorithm exist for predicting the Web QoE, e.g. Speed Index, Above the Fold, etc. Based on data from a measurement study, these algorithms shall be evaluated and compared in this work. In particular, a comparison of the metrics among the devices (Smartphone, Tablet, Desktop PC), which were used to generate the measurement data, is of interest.