Quantifying Mix Network Privacy Erosion with Generative Models

2025-06-10

AI Summary

Shuffling networks obscure metadata to enhance privacy, yet long-term privacy degradation remains difficult to quantify. This paper introduces the first dynamic privacy assessment framework based on generative large language models (LLMs) trained from scratch, modeling network traffic as a “language” and characterizing the progressive erosion of metadata privacy through iterative observations. Methodologically, it pioneers the application of generative modeling to shuffling privacy analysis; reveals the severe failure of conventional entropy and likelihood metrics under synthetic inference attacks; and discovers substantial privacy disparities among shuffling strategies with similar delays. Experiments demonstrate that the framework enables fine-grained quantification of privacy loss across diverse shuffling strategies, security parameters, and observation windows—achieving higher sample efficiency, stronger inference capability, and significantly improved assessment accuracy.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Security and Privacy: Large-scale security measurements
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems

Abstract

Modern mix networks improve over Tor and provide stronger privacy guarantees by robustly obfuscating metadata. As long as a message is routed through at least one honest mixnode, the privacy of the users involved is safeguarded. However, the complexity of the mixing mechanisms makes it difficult to estimate the cumulative privacy erosion occurring over time. This work uses a generative model trained on mixnet traffic to estimate the loss of privacy when users communicate persistently over a period of time. We train our large-language model from scratch on our specialized network traffic ``language'' and then use it to measure the sender-message unlinkability in various settings (e.g. mixing strategies, security parameters, observation window). Our findings reveal notable differences in privacy levels among mix strategies, even when they have similar mean latencies. In comparison, we demonstrate the limitations of traditional privacy metrics, such as entropy and log-likelihood, in fully capturing an adversary's potential to synthesize information from multiple observations. Finally, we show that larger models exhibit greater sample efficiency and superior capabilities implying that further advancements in transformers will consequently enhance the accuracy of model-based privacy estimates.

Problem

• Estimating cumulative privacy erosion in mix networks over time
• Measuring sender-message unlinkability under various mix strategies
• Evaluating limitations of traditional privacy metrics against adversaries

Innovation

• Generative model estimates mixnet privacy erosion
• Large-language model trained on network traffic
• Transformers enhance accuracy of privacy estimates