Paper 2023/199

MixFlow: Assessing Mixnets Anonymity with Contrastive Architectures and Semantic Network Information

Abstract

Traffic correlation attacks have illustrated challenges with protecting communication meta-data, yet short flows as in messaging applications like Signal have been protected by practical Mixnets such as Loopix from prior traffic correlation attacks. This paper introduces a novel traffic correlation attack against short-flow applications like Signal that are tunneled through practical Mixnets like Loopix. We propose the MixFlow model, an approach for analyzing the unlinkability of communications through Mix networks. As a prominent example, we do our analysis on Loopix. The MixFlow is a contrastive model that looks for semantic relationships between entry and exit flows, even if the traffic is tunneled through Mixnets that protect meta-data like Loopix via Poisson mixing delay and cover traffic. We use the MixFlow model to evaluate the resistance of Loopix Mix networks against an adversary that observes only the inflow and outflow of Mixnet and tries to correlate communication flows. Our experiments indicate that the MixFlow model is exceptionally proficient in connecting end-to-end flows, even when the Poison delay and cover traffic are increased. These findings challenge the conventional notion that adding Poisson mixing delay and cover traffic can obscure the metadata patterns and relationships between communicating parties. Despite the implementation of Poisson mixing countermeasures in Mixnets, MixFlow is still capable of effectively linking end-to-end flows, enabling the extraction of meta-information and correlation between inflows and outflows. Our findings have important implications for existing Poisson-mixing techniques and open up new opportunities for analyzing the anonymity and unlinkability of communication protocols.