When Mixnets Fail: Evaluating, Quantifying, and Mitigating the Impact of Adversarial Nodes in Mix Networks

currently deployed in practice-the Nym network [9] . In this design, nodes are arranged into layers such that each node in layer ℓ is only connected to nodes in layers ℓ-1 and ℓ+1. Consequently, each packet traverses exactly one node from each layer before reaching its destination. Additionally, for mixing purposes in Loopix, each mixnode flushes incoming packets after a random delay sampled from an exponential distribution [18], a suitable scheme for real-time communications. Within Loopix (Nym), client traffic is transmitted to its destination by fragmenting the traffic data into fixed-size packets using the Sphinx packet format [8], an efficient scheme that enhances packet unlinkability [25], [9] . Notably, each Sphinx packet carries a small payload (2 KB in Nym 1 ) and is routed independently via a path selected uniformly at random, consisting of one node from each mixnet layer. This design ensures that each route consistently carries a comparable number of packets, thereby reducing the GPA's ability to infer communication patterns based on traffic volume [9] . Although the aforementioned design choice effectively thwarts the threat posed by the GPA, its impact on the advantage of mixnode adversaries, compromising a subset of nodes within the mixnet, has not been thoroughly examined. Specifically, such adversaries can fully deanonymize clients if any of their packets traverse only adversarial nodes. While the threat of mixnode adversaries has been largely overlooked in state-of-the-art mixnet works [9], [16], [34], [35] , their lower infrastructure requirements make them a more practical threat in real-world scenarios. Design Goals. Given the practical threat of mixnode adversaries, this paper aims to: (1) precisely quantify the advantage gained by such adversaries in the Loopix (Nym) network; (2) introduce practical path assignment strategies designed to mitigate this advantage; (3) propose new evaluation metrics that accurately capture the adversarial threat; (4) conduct a thorough evaluation of our techniques and examine their impact on various aspects of mixnets, exploring whether reducing the advantage of mixnode adversaries can be achieved without significantly increasing the advantage of the GPA.