TrafficSliver: Fighting Website Fingerprinting Attacks with Traffic Splitting

Website fingerprinting (WFP) aims to infer information about the content of encrypted and anonymized connections by observing patterns of data flows based on the size and direction of packets. By collecting traffic traces at a malicious Tor entry node --- one of the weakest adversaries in the attacker model of Tor --- a passive eavesdropper can leverage the captured meta-data to reveal the websites visited by a Tor user. As recently shown, WFP is significantly more effective and realistic than assumed. Concurrently, former WFP defenses are either infeasible for deployment in real-world settings or defend against specific WFP attacks only. To limit the exposure of Tor users to WFP, we propose novel lightweight WFP defenses, TrafficSliver, which successfully counter today's WFP classifiers with reasonable bandwidth and latency overheads and, thus, make them attractive candidates for adoption in Tor. Through user-controlled splitting of traffic over multiple Tor entry nodes, TrafficSliver limits the data a single entry node can observe and distorts repeatable traffic patterns exploited by WFP attacks. We first propose a network-layer defense, in which we apply the concept of multipathing entirely within the Tor network. We show that our network-layer defense reduces the accuracy from more than 98% to less than 16% for all state-of-the-art WFP attacks without adding any artificial delays or dummy traffic. We further suggest an elegant client-side application-layer defense, which is independent of the underlying anonymization network. By sending single HTTP requests for different web objects over distinct Tor entry nodes, our application-layer defense reduces the detection rate of WFP classifiers by almost 50 percentage points. Although it offers lower protection than our network-layer defense, it provides a security boost at the cost of a very low implementation overhead and is fully compatible with today's Tor network.