Procjena performansi usluge YouTube na platformi iOS na osnovu analize kriptiranog prometa uporabom tehnike strojnog učenja

master's thesis

• Title (english): A Machine Learning Approach to Estimating YouTube Performance on the iOS Platform Based on the Analysis of Encrypted Network Traffic
• Author: Petra Rebernjak
• Mentor: Lea Skorin-Kapov (mentor)
• Committee member: Lea Skorin-Kapov (predsjednik povjerenstva)
• Committee member: Ognjen Dobrijević (član povjerenstva)
• Committee member: Mirko Sužnjević (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2018-07-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

YouTube je danas jedna od najpopularnijih usluga za dostavu video sadržaja na Internetu, koja se temelji na prilagodljivom strujanju putem protokola HTTP. Davatelji mrežnih usluga uglavnom nemaju uvid u performanse i parametre kvalitete YouTube tokova koji prolaze njihovom mrežom, obzirom da je YouTube-ov promet kriptiran. Rješenja koja koriste tehnike strojnog učenja za procjenu iskustvene kvalitete temeljem analize značajki mrežnog prometa mogla bi riješiti ovaj problem.

S ciljem analize ponašanja YouTube-a na operacijskom sustavu iOS i razvoja metodologije za procjenu iskustvene kvalitete korisnika isključivo na temelju statistika šifriranog mrežnog prometa, u testnim eksperimentima prikupljani su podatci aplikacijskog sloja te je sniman promet u mreži. Podatci aplikacijskog sloja bilježeni su na temelju "Stats for Nerds" podataka dostupnih unutar YouTube aplikacije. Svaki video iz testnog eksperimenta predstavljen je značajkama iz prometa te označen stvarnim razredom iskustvene kvalitete na temelju podataka aplikacijskog sloja. Takvi podatci korišteni su za treniranje i evaluaciju modela strojnog učenja.

Konačni skup podataka sastoji se od 329 YouTube videa označenih "low", "medium" i "high" razredima iskustvene kvalitete. Nekoliko je klasifikacijskih modela temeljenih na decizijskim stablima trenirano i evaluirano. F1-mjera najboljeg modela postiže 96.4% na prikupljenom skupu podataka.

Abstract (english)

YouTube relies on HTTP adaptive streaming and is one of the most popular media delivery services on the Internet today. Currently, YouTube traffic transmitted over the network is encrypted, hence network providers are missing mechanisms for estimating application performance degradation events, such as stalling and quality switches. Solutions based on machine learning approaches for estimating performance solely from encrypted network traffic show promising results.

In order to analyse the YouTube adaptation algorithm on the iOS platform, and develop a methodology for estimating QoE based on encrypted traffic analysis, both application-layer events, derived from the native YouTube player by enabling "Stats for Nerds" option, and encrypted network traffic were collected during experiments. Input to machine learning models is prepared by labelling encrypted network traffic features with QoE class derived from the application-level KPIs for each video played in the experiments.

A total of 329 YouTube videos were streamed over 33 bandwidth scenarios, and each video was labeled with "low", "medium" or "high" QoE class. Multiple tree-based classification models were trained and evaluated. Overall performance of the developed machine learning model reached the F1-score of 96.4% on the collected dataset.

• Keywords: iskustvena kvaliteta korisnika; kriptirani mrežni promet; prilagodljivo strujanje putem protokola HTTP; YouTube; strojno učenje
• Keywords (english): Quality of Experience; Encrypted Network Traffic; Dynamic Adaptive Streaming over HTTP; YouTube; Machine Learning
• Language: english
• URN:NBN: urn:nbn:hr:168:614126
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-03-14 00:24:44