Optimization of load balancing in anonymous dynamical networks with application to Tor

Darir, Hussein

Permalink

https://hdl.handle.net/2142/120326 Copy

Description

Title

Optimization of load balancing in anonymous dynamical networks with application to Tor

Author(s)

Darir, Hussein

Issue Date

2022-12-15

Director of Research (if dissertation) or Advisor (if thesis)

• Dullerud, Geir
• Borisov, Nikita

Doctoral Committee Chair(s)

• Dullerud, Geir
• Borisov, Nikita

Committee Member(s)

• Mitra, Sayan
• Mehta, Prashant

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Optimization
• load balancing
• Tor network
• maximum likelihood estimation
• probabilistic program
• probabilistic model
• Python
• Shadow simulator
• estimation

Abstract

This thesis is about cyber-physical systems that play an essential part in connecting the physical world to the world of communication and computation at the time of writing. The application of cyber-physical technology is of particular importance in the Internet where the anonymity of users is an important property that should be addressed. Tor is a system that helps protect online privacy and circumvent any censorship that may be present on the Internet. It has millions of daily active users. Tor uses a network of volunteer relays to help encrypt users’ traffic and obscure its source and destination. Each Tor user requires three nodes for their traffic to go through. Those relays should be chosen in a way so that no node is overloaded and hence no disparity is created between the service offered to each client. In this work, the goal is to better understand the dynamics of bandwidth measurement and path allocation in Tor and design an improved measurement scheme. To this end, a mathematical model of bandwidth measurement in the Tor network is derived. This model makes few simplifying assumptions but allow us to model the behavior of estimation algorithms. None of the previous work on relay capacity estimation developed a probabilistic model of measurements in Tor. Maximum likelihood is applied using this model to estimate the capacities of relays in the network, this algorithm is called MLEFlow. Furthermore, analytical bounds and guarantees of convergence for the estimates are derived to show the higher accuracy of the new proposed algorithm. Additionally, in this work, the currently deployed algorithm is shown to be equivalent to a one step maximum likelihood estimation, laying the mathematical foundation behind its implementation. A more complex, and hence more realistic, probabilistic model of the Tor network is then developed. To be able to overcome the analytical barrier of solving the estimation problem using this complex model, Probabilistic Programming is used to estimate the capacities of relays. This algorithm is called ProbFlow. A probabilistic program is implemented in Pyro, a probabilistic programming language in Python. Two different ways of simulating the network are used to show the benefits of using the proposed algorithms. First, a custom-built flow-based simulator, written in Python, simulates the basic behavior of the entire Tor network. The second simulation platform used is the Shadow simulator, the state-of-art simulator of the Tor network providing high-fidelity simulation of the network. All the proposed algorithms are shown to result in much more accurate estimation of network capacities of relays, which, in turn, result in significantly better balancing of load for user traffic.

Graduation Semester

2023-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Hussein Darir

Owning Collections