The path less travelled: Overcoming Tor's bottlenecks with traffic splitting

Tor is the most popular low-latency anonymity network for enhancing ordinary users' online privacy and resisting censorship. While it has grown in popularity, Tor has a variety of performance problems that result in poor quality of service, a strong disincentive to use the system, and weaker anonymity properties for all users. We observe that one reason why Tor is slow is due to low-bandwidth volunteer-operated routers. When clients use a low-bandwidth router, their throughput is limited by the capacity of the slowest node. With the introduction of bridges - unadvertised Tor routers that provide Tor access to users within censored regimes like China - low-bandwidth Tor routers are becoming more common and essential to Tor's ability to resist censorship. In this paper, we present Conflux, a dynamic traffic-splitting approach that assigns traffic to an overlay path based on its measured latency. Because it enhances the load-balancing properties of the network, Conflux considerably increases performance for clients using low-bandwidth bridges. Moreover, Conflux significantly improves the experience of users who watch streaming videos online. Through live measurements and a whole-network evaluation conducted on a scalable network emulator, we show that our approach offers an improvement of approximately 30% in expected download time for web browsers who use Tor bridges and for streaming application users. We also show that Conflux introduces only slight tradeoffs between users' anonymity and performance.