Quantifying Mix Network Privacy Erosion with Generative Models

Vasilios Mavroudis, Tariq ElahiarXiv 2025 · DOI · arXiv

Abstract

Mix networks (mixnets) are designed to provide strong metadata anonymity by shuffling and delaying messages. However, persistent communication patterns can lead to privacy erosion over time, a phenomenon that is difficult to quantify using traditional static metrics. This paper presents LLMix, a generative AI framework that treats mixnet traffic as a sequence-based language. By training transformer models (Longformers) on network traces, we demonstrate that a sophisticated adversary can achieve high sender identification accuracy (up to 95.8%) even when traditional metrics like entropy suggest high levels of anonymity. Our findings highlight a significant "statistical gap" in current privacy measurement techniques and suggest that mixnet designers must account for cumulative information leakage.

Introduction

Mix networks, such as Nym and HOPR, offer enhanced privacy compared to onion routing by decoupling the timing and order of incoming and outgoing messages. While theoretically robust, these systems face the challenge of privacy erosion: the gradual accumulation of information by an observer over multiple communication rounds. Traditional privacy metrics, such as entropy or likelihood difference, often provide "snapshots" of a single message's anonymity but fail to capture the cumulative patterns that a global passive adversary might exploit over days or weeks.

Methodology

The LLMix framework shifts privacy estimation from manual feature engineering to automated generative modeling. The approach consists of three primary components:

1. Network Traffic Encoding

Mixnet activity is represented as a sequence of numerical identifiers corresponding to network links. This transforms raw traffic into a formal language where each "word" (link ID) has a precise meaning, eliminating lexical ambiguity. This encoding allows the model to learn the "grammar" of message movement through the network.

Figure 1: Network traffic encoding for single and multi-node mixnet configurations.

2. Generative Modeling

The researchers trained a Longformer architecture from scratch on these encoded sequences. Unlike pre-trained models, this model learns the specific mixing strategies, including Threshold, Pool, and Poisson mixing. Curriculum learning was employed to stabilize training, starting with simple configurations before moving to complex multi-node topologies.

3. Privacy Games

To quantify privacy, the problem is framed as a Sender-Message Unlinkability game. In this setup, a "Challenger" presents an adversary with a recipient and two potential senders. The adversary must identify the correct sender based on the observed traffic traces. This framework allows for a direct measurement of the adversary's success rate as a proxy for privacy leakage.

Figure 4: Experimental setup for the Sender-Message Unlinkability privacy game.

Results and Discussion

The study utilized a custom Python-based mixnet testbed with 100 active users. Key findings include:

  • Cumulative Leakage: Privacy is not a binary state but a degrading resource. While a single message might be well-hidden (58% accuracy), identification accuracy climbs to nearly 96% after approximately 20 messages (4096 observations).
  • The Statistical Gap: A critical discovery was the failure of traditional metrics. In one scenario, entropy suggested the sender was indistinguishable among 57 users, yet the LLMix model identified the sender with high confidence. This indicates that industry-standard metrics are dangerously optimistic.
  • Mixing Strategy Trade-offs: Poisson mixing, while beneficial for reducing message latency, was found to be less robust than batch-based threshold mixing for the same average latency. Furthermore, adding a "pool" to a threshold mix provided superior privacy compared to simply increasing the threshold.

Conclusion

As generative models become more capable of synthesizing complex patterns, the tools used to measure privacy must keep pace. LLMix demonstrates that transformers can "read" network traffic to deanonymize users far more effectively than previously thought. Future work will expand this framework to include active adversaries and non-uniform traffic patterns, providing a more comprehensive tool for mixnet security audits.

Summary updated

BibTeX
@misc{mavroudis2025quantifying,
  author = {Vasilios Mavroudis and Tariq Elahi},
  title = {{Quantifying Mix Network Privacy Erosion with Generative Models}},
  year = {2025},
  eprint = {2506.08918},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.08918},
}
Figure