Nikolas Wehner - Lehrstuhl für Informatik III

Telefon	(0931) 31-88150
Telefax	(0931) 31-86632
E-Mail	nikolas.wehner@informatik.uni-wuerzburg.de
Raum	A208
Anschrift	Lehrstuhl für Informatik III Am Hubland D-97074 Würzburg

Contents: Overview | Publications | Teaching | Working Group | Projects

Research Interests

My research interests include the following topics:

Quality of Experience (QoE) of Internet applications
Measurement and analytics of (encrypted) network traffic
Artificial intelligence and machine learning for networks
Performance analysis and modeling of communication systems

Recent Publications

Fingerprinting Webpages and Smartphone Apps from Encrypted Network Traffic with WebScanner Casas, Pedro; Wehner, Nikolas; Wassermann, Sarah; Seufert, Michael in Global Internet (GI) Symposium (2022).
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
The wide adoption of end-to-end encryption drastically limits the visibility Internet Service Providers (ISPs) have on the services consumed by their customers. This is particularly challenging for monitoring and analysis of web and apps traffic, which is highly complex and heterogeneous. Loading a web page or app today requires tens of flows to download the various resources located in distributed cloud servers from different content providers. We introduce WebScanner, a webpage and app fingerprinting approach capable to identify all the traffic flows corresponding to individual webpage and app loading sessions within concurrent web pages traffic, enabling highly detailed, per webpage analysis in practical deployments. Different from the state of the art in web and app traffic fingerprinting, and as our most relevant contribution, WebScanner automatically performs the parsing of all the traffic generated by a web visit and its isolation from concurrent traffic, instead of assuming that an external oracle system does so. WebScanner also implements a deep fingerprinting approach to detect user action-dependent traffic from apps, relying on simple machine learning models and strong input features as fingerprints. Extensive evaluation across a large measurement dataset of popular webpages and mobile apps confirms the outstanding performance of WebScanner, identifying the top-500 Alexa websites with precision and recall (P/R) above 95%, isolating their full contents with P/R above 80% for up to 15 concurrent webpages, and detecting specific action-dependent apps traffic with average P/R above 92%.

@inproceedings{info3-inproceedings-2022-18, abstract = {The wide adoption of end-to-end encryption drastically limits the visibility Internet Service Providers (ISPs) have on the services consumed by their customers. This is particularly challenging for monitoring and analysis of web and apps traffic, which is highly complex and heterogeneous. Loading a web page or app today requires tens of flows to download the various resources located in distributed cloud servers from different content providers. We introduce WebScanner, a webpage and app fingerprinting approach capable to identify all the traffic flows corresponding to individual webpage and app loading sessions within concurrent web pages traffic, enabling highly detailed, per webpage analysis in practical deployments. Different from the state of the art in web and app traffic fingerprinting, and as our most relevant contribution, WebScanner automatically performs the parsing of all the traffic generated by a web visit and its isolation from concurrent traffic, instead of assuming that an external oracle system does so. WebScanner also implements a deep fingerprinting approach to detect user action-dependent traffic from apps, relying on simple machine learning models and strong input features as fingerprints. Extensive evaluation across a large measurement dataset of popular webpages and mobile apps confirms the outstanding performance of WebScanner, identifying the top-500 Alexa websites with precision and recall (P/R) above 95%, isolating their full contents with P/R above 80% for up to 15 concurrent webpages, and detecting specific action-dependent apps traffic with average P/R above 92%.}, address = {Paris, France}, author = {Casas, Pedro and Wehner, Nikolas and Wassermann, Sarah and Seufert, Michael}, booktitle = {Global Internet (GI) Symposium}, keywords = {ucn}, month = 11, title = {Fingerprinting Webpages and Smartphone Apps from Encrypted Network Traffic with WebScanner}, year = 2022 }

%0 Conference Paper %1 info3-inproceedings-2022-18 %A Casas, Pedro %A Wehner, Nikolas %A Wassermann, Sarah %A Seufert, Michael %B Global Internet (GI) Symposium %C Paris, France %D 2022 %T Fingerprinting Webpages and Smartphone Apps from Encrypted Network Traffic with WebScanner %X The wide adoption of end-to-end encryption drastically limits the visibility Internet Service Providers (ISPs) have on the services consumed by their customers. This is particularly challenging for monitoring and analysis of web and apps traffic, which is highly complex and heterogeneous. Loading a web page or app today requires tens of flows to download the various resources located in distributed cloud servers from different content providers. We introduce WebScanner, a webpage and app fingerprinting approach capable to identify all the traffic flows corresponding to individual webpage and app loading sessions within concurrent web pages traffic, enabling highly detailed, per webpage analysis in practical deployments. Different from the state of the art in web and app traffic fingerprinting, and as our most relevant contribution, WebScanner automatically performs the parsing of all the traffic generated by a web visit and its isolation from concurrent traffic, instead of assuming that an external oracle system does so. WebScanner also implements a deep fingerprinting approach to detect user action-dependent traffic from apps, relying on simple machine learning models and strong input features as fingerprints. Extensive evaluation across a large measurement dataset of popular webpages and mobile apps confirms the outstanding performance of WebScanner, identifying the top-500 Alexa websites with precision and recall (P/R) above 95%, isolating their full contents with P/R above 80% for up to 15 concurrent webpages, and detecting specific action-dependent apps traffic with average P/R above 92%.
A Vital Improvement? Relating Google’s Core Web Vitals to Actual Web QoE Wehner, Nikolas; Amir, Monisha; Seufert, Michael; Schatz, Raimund; Hoßfeld, Tobias in 14th International Conference on Quality of Multimedia Experience (QoMEX), Best Paper Award (2022).
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Providing sophisticated Web Quality of Experience (QoE) has become paramount for web service providers and network operators alike. Due to advances in Web technologies (HTML5, responsive design, etc.), traditional Web QoE models focusing mainly on loading times have to be refined and improved. In this work, we relate Google's Core Web Vitals, a set of metrics for improving user experience, to the loading time aspects of Web QoE. To this end, we first perform objective measurements in the Web using Google's Lighthouse. To close the gap between metrics and experience, we complement these objective measurements with subjective assessment by performing multiple crowdsourcing QoE studies. In these studies, we use CWeQS, a customized framework to emulate the entire web page loading process, and ask users for their experience while controlling the Core Web Vitals. Our results suggest that the Core Web Vitals have less predictive value for Web QoE than expected and that page loading times remain the main influence factor in this context.

@inproceedings{info3-inproceedings-2022-12, abstract = {Providing sophisticated Web Quality of Experience (QoE) has become paramount for web service providers and network operators alike. Due to advances in Web technologies (HTML5, responsive design, etc.), traditional Web QoE models focusing mainly on loading times have to be refined and improved. In this work, we relate Google's Core Web Vitals, a set of metrics for improving user experience, to the loading time aspects of Web QoE. To this end, we first perform objective measurements in the Web using Google's Lighthouse. To close the gap between metrics and experience, we complement these objective measurements with subjective assessment by performing multiple crowdsourcing QoE studies. In these studies, we use CWeQS, a customized framework to emulate the entire web page loading process, and ask users for their experience while controlling the Core Web Vitals. Our results suggest that the Core Web Vitals have less predictive value for Web QoE than expected and that page loading times remain the main influence factor in this context.}, address = {Lippstadt, Germany}, author = {Wehner, Nikolas and Amir, Monisha and Seufert, Michael and Schatz, Raimund and Hoßfeld, Tobias}, booktitle = {14th International Conference on Quality of Multimedia Experience (QoMEX), Best Paper Award}, keywords = {ucn}, month = 9, title = {A Vital Improvement? Relating Google's Core Web Vitals to Actual Web QoE}, year = 2022 }

%0 Conference Paper %1 info3-inproceedings-2022-12 %A Wehner, Nikolas %A Amir, Monisha %A Seufert, Michael %A Schatz, Raimund %A Hoßfeld, Tobias %B 14th International Conference on Quality of Multimedia Experience (QoMEX), Best Paper Award %C Lippstadt, Germany %D 2022 %T A Vital Improvement? Relating Google's Core Web Vitals to Actual Web QoE %X Providing sophisticated Web Quality of Experience (QoE) has become paramount for web service providers and network operators alike. Due to advances in Web technologies (HTML5, responsive design, etc.), traditional Web QoE models focusing mainly on loading times have to be refined and improved. In this work, we relate Google's Core Web Vitals, a set of metrics for improving user experience, to the loading time aspects of Web QoE. To this end, we first perform objective measurements in the Web using Google's Lighthouse. To close the gap between metrics and experience, we complement these objective measurements with subjective assessment by performing multiple crowdsourcing QoE studies. In these studies, we use CWeQS, a customized framework to emulate the entire web page loading process, and ask users for their experience while controlling the Core Web Vitals. Our results suggest that the Core Web Vitals have less predictive value for Web QoE than expected and that page loading times remain the main influence factor in this context.
X-Ray Goggles for the ISP: Improving in-Network Web and App QoE Monitoring with Deep Learning Casas, Pedro; Wassermann, Sarah; Seufert, Michael; Wehner, Nikolas; Dinica, Olivia; Hoßfeld, Tobias in Network Traffic Measurement and Analysis Conference (TMA) (2022).
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
The wide adoption of end-to-end encryption is drastically limiting the visibility Internet Service Providers (ISPs) have on the performance of the services consumed by their customers. In times of strong competition, where customer experience plays a key role in churn management, ISPs require novel solutions enabling network-wide Quality of Experience (QoE) monitoring. To this end, we present DeepQoE, a deep-learning based approach to infer the QoE of web services and mobile applications from the ISP perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived from the encrypted stream of bytes as input to deep Convolutional Neural Networks (CNNs), DeepQoE infers the Speed Index of web browsing sessions and general mobile apps with unprecedented accuracy, improving the state of the art by more than 25%, and reducing the QoE inference error in terms of mean opinion scores by nearly 40%. DeepQoE implements a web fingerprinting solution to identify individual web browsing sessions within concurrent web pages traffic, enabling highly detailed, per web page QoE inference in practical deployments. Extensive evaluations over a large and heterogeneous dataset composed of web and app measurements, using different device types and for top-popular websites and apps, confirm the out-performance of DeepQoE over previously used shallow-learning models, as well as the deep-model generalization to different devices, web pages, apps, and network setups. DeepQoE is the first deployable system providing such a deep, highly-detailed QoE for individual web browsing and mobile apps over encrypted traffic, using deep learning models on heterogeneous measurements.

@inproceedings{info3-inproceedings-2022-11, abstract = {The wide adoption of end-to-end encryption is drastically limiting the visibility Internet Service Providers (ISPs) have on the performance of the services consumed by their customers. In times of strong competition, where customer experience plays a key role in churn management, ISPs require novel solutions enabling network-wide Quality of Experience (QoE) monitoring. To this end, we present DeepQoE, a deep-learning based approach to infer the QoE of web services and mobile applications from the ISP perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived from the encrypted stream of bytes as input to deep Convolutional Neural Networks (CNNs), DeepQoE infers the Speed Index of web browsing sessions and general mobile apps with unprecedented accuracy, improving the state of the art by more than 25%, and reducing the QoE inference error in terms of mean opinion scores by nearly 40%. DeepQoE implements a web fingerprinting solution to identify individual web browsing sessions within concurrent web pages traffic, enabling highly detailed, per web page QoE inference in practical deployments. Extensive evaluations over a large and heterogeneous dataset composed of web and app measurements, using different device types and for top-popular websites and apps, confirm the out-performance of DeepQoE over previously used shallow-learning models, as well as the deep-model generalization to different devices, web pages, apps, and network setups. DeepQoE is the first deployable system providing such a deep, highly-detailed QoE for individual web browsing and mobile apps over encrypted traffic, using deep learning models on heterogeneous measurements.}, author = {Casas, Pedro and Wassermann, Sarah and Seufert, Michael and Wehner, Nikolas and Dinica, Olivia and Hoßfeld, Tobias}, booktitle = {Network Traffic Measurement and Analysis Conference (TMA)}, keywords = {sys:hidden:inproceedings}, month = 6, title = {X-Ray Goggles for the ISP: Improving in-Network Web and App QoE Monitoring with Deep Learning}, year = 2022 }

%0 Conference Paper %1 info3-inproceedings-2022-11 %A Casas, Pedro %A Wassermann, Sarah %A Seufert, Michael %A Wehner, Nikolas %A Dinica, Olivia %A Hoßfeld, Tobias %B Network Traffic Measurement and Analysis Conference (TMA) %D 2022 %T X-Ray Goggles for the ISP: Improving in-Network Web and App QoE Monitoring with Deep Learning %X The wide adoption of end-to-end encryption is drastically limiting the visibility Internet Service Providers (ISPs) have on the performance of the services consumed by their customers. In times of strong competition, where customer experience plays a key role in churn management, ISPs require novel solutions enabling network-wide Quality of Experience (QoE) monitoring. To this end, we present DeepQoE, a deep-learning based approach to infer the QoE of web services and mobile applications from the ISP perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived from the encrypted stream of bytes as input to deep Convolutional Neural Networks (CNNs), DeepQoE infers the Speed Index of web browsing sessions and general mobile apps with unprecedented accuracy, improving the state of the art by more than 25%, and reducing the QoE inference error in terms of mean opinion scores by nearly 40%. DeepQoE implements a web fingerprinting solution to identify individual web browsing sessions within concurrent web pages traffic, enabling highly detailed, per web page QoE inference in practical deployments. Extensive evaluations over a large and heterogeneous dataset composed of web and app measurements, using different device types and for top-popular websites and apps, confirm the out-performance of DeepQoE over previously used shallow-learning models, as well as the deep-model generalization to different devices, web pages, apps, and network setups. DeepQoE is the first deployable system providing such a deep, highly-detailed QoE for individual web browsing and mobile apps over encrypted traffic, using deep learning models on heterogeneous measurements.
Not all Web Pages are Born the Same. Content Tailored Learning for Web QoE Inference Casas, Pedro; Wassermann, Sarah; Wehner, Nikolas; Seufert, Michael; Hoßfeld, Tobias in IEEE International Symposium on Measurements & Networking (M&N 2022) (2022).
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Web Quality of Experience (QoE) monitoring is a critical task for Internet Service Providers (ISPs), especially due to the key role played by customer experience in churn management. Previously, we have tackled the problem of Web QoE inference from the ISP perspective, relying on passive measurement of encrypted network traffic and machine learning models. In this paper, we exploit the broad heterogeneity of contents embedded in web pages to improve the state of the art performance in Web QoE inference, relying on web-content learning model tailoring. By analyzing the top-500 most popular web pages of the Internet through unsupervised learning, we discover different web page content classes which realize significantly different Web QoE inference performance. We train supervised learning inference models separately for each of these classes, using the well-known Speed Index (SI) metric as proxy to Web QoE. Empirical evaluations on a large corpus of Web QoE measurements for top popular websites demonstrate that our combined content-tailored approach improves the inference performance of the SI by almost 30% with respect to previous single-model approaches, reducing the QoE inference error in terms of mean opinion scores by more than 40%.

@inproceedings{info3-inproceedings-2022-10, abstract = {Web Quality of Experience (QoE) monitoring is a critical task for Internet Service Providers (ISPs), especially due to the key role played by customer experience in churn management. Previously, we have tackled the problem of Web QoE inference from the ISP perspective, relying on passive measurement of encrypted network traffic and machine learning models. In this paper, we exploit the broad heterogeneity of contents embedded in web pages to improve the state of the art performance in Web QoE inference, relying on web-content learning model tailoring. By analyzing the top-500 most popular web pages of the Internet through unsupervised learning, we discover different web page content classes which realize significantly different Web QoE inference performance. We train supervised learning inference models separately for each of these classes, using the well-known Speed Index (SI) metric as proxy to Web QoE. Empirical evaluations on a large corpus of Web QoE measurements for top popular websites demonstrate that our combined content-tailored approach improves the inference performance of the SI by almost 30% with respect to previous single-model approaches, reducing the QoE inference error in terms of mean opinion scores by more than 40%.}, author = {Casas, Pedro and Wassermann, Sarah and Wehner, Nikolas and Seufert, Michael and Hoßfeld, Tobias}, booktitle = {IEEE International Symposium on Measurements & Networking (M&N 2022)}, keywords = {sys:hidden:inproceedings}, month = 7, title = {Not all Web Pages are Born the Same. Content Tailored Learning for Web QoE Inference}, year = 2022 }

%0 Conference Paper %1 info3-inproceedings-2022-10 %A Casas, Pedro %A Wassermann, Sarah %A Wehner, Nikolas %A Seufert, Michael %A Hoßfeld, Tobias %B IEEE International Symposium on Measurements & Networking (M&N 2022) %D 2022 %T Not all Web Pages are Born the Same. Content Tailored Learning for Web QoE Inference %X Web Quality of Experience (QoE) monitoring is a critical task for Internet Service Providers (ISPs), especially due to the key role played by customer experience in churn management. Previously, we have tackled the problem of Web QoE inference from the ISP perspective, relying on passive measurement of encrypted network traffic and machine learning models. In this paper, we exploit the broad heterogeneity of contents embedded in web pages to improve the state of the art performance in Web QoE inference, relying on web-content learning model tailoring. By analyzing the top-500 most popular web pages of the Internet through unsupervised learning, we discover different web page content classes which realize significantly different Web QoE inference performance. We train supervised learning inference models separately for each of these classes, using the well-known Speed Index (SI) metric as proxy to Web QoE. Empirical evaluations on a large corpus of Web QoE measurements for top popular websites demonstrate that our combined content-tailored approach improves the inference performance of the SI by almost 30% with respect to previous single-model approaches, reducing the QoE inference error in terms of mean opinion scores by more than 40%.
DeepCrypt - Deep Learning for QoE Monitoring and Fingerprinting of User Actions in Adaptive Video Streaming Casas, Pedro; Seufert, Michael; Wassermann, Sarah; Gardlo, Bruno; Wehner, Nikolas; Schatz, Raimund in IEEE Conference on Network Softwarization (IEEE NetSoft 2022) (2022).
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
We introduce DeepCrypt, a deep-learning based approach to analyze YouTube video streaming Quality of Experience (QoE) from the Internet Service Provider (ISP) perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived on-line from the encrypted stream of bytes, DeepCrypt can infer six different video QoE indicators capturing the user-perceived performance of the service, including the initial playback delay, the number and frequency of rebuffering events, the video playback quality, the video encoding bitrate, and the number of quality changes. DeepCrypt offers deep visibility into the behavior of the end-user, enabling the fingerprinting and detection of different user actions on the video player, such as video pauses and playback scrubbing (forward, backward, out-of-buffer), offering a complete visibility on the video streaming process from in-network traffic measurements. Extensive evaluations over a large and heterogeneous dataset composed of mobile and fixed-line measurements, using the YouTube HTML5 player, the native YouTube mobile app, as well as a generic HTML5 video player built on top of open source libraries, and considering measurements collected at different ISPs, confirm the out-performance of DeepCrypt over previously used shallow-learning models, as well as the deep-model generalization to different video players and network setups. To the best of our knowledge, this is the first paper tackling such a combined QoE and user-actions approach for video streaming applications, using deep learning models over heterogeneous measurements.

@inproceedings{info3-inproceedings-2022-9, abstract = {We introduce DeepCrypt, a deep-learning based approach to analyze YouTube video streaming Quality of Experience (QoE) from the Internet Service Provider (ISP) perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived on-line from the encrypted stream of bytes, DeepCrypt can infer six different video QoE indicators capturing the user-perceived performance of the service, including the initial playback delay, the number and frequency of rebuffering events, the video playback quality, the video encoding bitrate, and the number of quality changes. DeepCrypt offers deep visibility into the behavior of the end-user, enabling the fingerprinting and detection of different user actions on the video player, such as video pauses and playback scrubbing (forward, backward, out-of-buffer), offering a complete visibility on the video streaming process from in-network traffic measurements. Extensive evaluations over a large and heterogeneous dataset composed of mobile and fixed-line measurements, using the YouTube HTML5 player, the native YouTube mobile app, as well as a generic HTML5 video player built on top of open source libraries, and considering measurements collected at different ISPs, confirm the out-performance of DeepCrypt over previously used shallow-learning models, as well as the deep-model generalization to different video players and network setups. To the best of our knowledge, this is the first paper tackling such a combined QoE and user-actions approach for video streaming applications, using deep learning models over heterogeneous measurements.}, author = {Casas, Pedro and Seufert, Michael and Wassermann, Sarah and Gardlo, Bruno and Wehner, Nikolas and Schatz, Raimund}, booktitle = {IEEE Conference on Network Softwarization (IEEE NetSoft 2022)}, keywords = {sys:hidden:inproceedings}, month = 6, title = {DeepCrypt - Deep Learning for QoE Monitoring and Fingerprinting of User Actions in Adaptive Video Streaming}, year = 2022 }

%0 Conference Paper %1 info3-inproceedings-2022-9 %A Casas, Pedro %A Seufert, Michael %A Wassermann, Sarah %A Gardlo, Bruno %A Wehner, Nikolas %A Schatz, Raimund %B IEEE Conference on Network Softwarization (IEEE NetSoft 2022) %D 2022 %T DeepCrypt - Deep Learning for QoE Monitoring and Fingerprinting of User Actions in Adaptive Video Streaming %X We introduce DeepCrypt, a deep-learning based approach to analyze YouTube video streaming Quality of Experience (QoE) from the Internet Service Provider (ISP) perspective, relying exclusively on the analysis of encrypted network traffic. Using raw features derived on-line from the encrypted stream of bytes, DeepCrypt can infer six different video QoE indicators capturing the user-perceived performance of the service, including the initial playback delay, the number and frequency of rebuffering events, the video playback quality, the video encoding bitrate, and the number of quality changes. DeepCrypt offers deep visibility into the behavior of the end-user, enabling the fingerprinting and detection of different user actions on the video player, such as video pauses and playback scrubbing (forward, backward, out-of-buffer), offering a complete visibility on the video streaming process from in-network traffic measurements. Extensive evaluations over a large and heterogeneous dataset composed of mobile and fixed-line measurements, using the YouTube HTML5 player, the native YouTube mobile app, as well as a generic HTML5 video player built on top of open source libraries, and considering measurements collected at different ISPs, confirm the out-performance of DeepCrypt over previously used shallow-learning models, as well as the deep-model generalization to different video players and network setups. To the best of our knowledge, this is the first paper tackling such a combined QoE and user-actions approach for video streaming applications, using deep learning models over heterogeneous measurements.

A comprehensive list of all publications can be found here: Publications

Nikolas Wehner M. Sc.

Research Interests

Recent Publications

Hinweis zum Datenschutz

Hinweis zum Datenschutz

Bildnachweise