Measurements Of A Latency-biased Expander Topology In The Tor Anonymity System

Abstract

Anonymous communication systems protect the privacy of their users by hiding who is communicating with whom. With the widespread use of the Internet, anonymity systems are all the more essential to support applications having strong privacy requirements such as intelligence gathering, military communications, or e-voting protocols. Anonymity systems must balance security and performance to remain popular with their users. In this work, we perform measurements on anonymity systems to improve their performance. We use the Vivaldi network coordinate system to efficiently map out the relative delays between hosts. Using this data, we create an overlay expander network topology that is biased to use lower latency links instead of randomly selecting nodes. Our primary contribution is the design and execution of a set of experiments to evaluate the performance of this approach. These experiments are performed using a private deployment of Tor, a popular anonymity system, running on PlanetLab, a globally distributed testbed. Our testbed is comprised of 100 Tor relay nodes, five trusted directory servers and 10 geographically distributed clients, with each of the relays running a common implementation of Vivaldi to compute its virtual coordinates and reporting the same to a trusted directory server. The directory server uses this information to construct an expander graph topology with a bias towards faster links. We show that when the network topology is created with a bias towards lower latency edges, there is a significant improvement in performance compared to using random links on our topology.