Enⅽrypteⅾ Traffiⅽ Anaⅼysis through Expⅼainabⅼe Ⅿaⅽhine Ⅼearning

Please enable javascript in your browser.

Moghaddas Esfahani, Mohammad Reza | 2022

135 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 55669 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Jalili, Rasool
Abstract:
Impressive progress in hardwares and developing encryption algorithms in last two decades are caused increase in using encryption protocols in network communications. In last decade, users use privacy preserving networks like Jap and Tor to protect their privacy. These networks protect users' data from eavesdroppers by using three-layer encryption and intermediate nodes between user and target website. Recent researches show that Deep Neural Networks can predict websites viewed by users with high accuracy. In other words, privacy preserving networks suffer from information leakage. In this research, we introduced some of the most powerful methods in encrypted traffic classification and then trained an explainable-by-design model uses modified deep fingerprinting model's convolutional neural network as feature extractor, then we illustrated how to analyze this model's explanations. In the end, we explained how to assess this model and then assess model classification power and explainability quality. for classification power assessment, 'Accuracy' is a good choice for our dataset and for explainability quality we have used 'Robustness' and 'Effectiveness' metrics
Keywords:
Traffic Analysis ; Encrypted Traffic ; Explainable Machine Learning ; Model Transparency ; Interpreter Model ; Privacy Preserving