One-Prompt Censorship Evasion via Generative Diffusion Models

2026-06-21

AI Summary

Existing censorship circumvention tools often suffer from limited robustness and complex configuration when confronted with diverse deep learning–based censorship systems. This work reframes network censorship evasion as a semantic-level image-to-image editing task, leveraging the “world knowledge” embedded in large diffusion models to automatically transform sensitive traffic into benign patterns through a single natural language instruction. Built upon an instruction-tuned generative diffusion architecture, the proposed method enables one-click, cross-modal semantic editing of network traffic without manual parameter tuning or domain-specific scripting. Experimental results demonstrate that it significantly outperforms existing baselines on both industrial-grade rule-based middleware and learning-based classifiers. Notably, merely altering the natural language prompt allows the approach to effectively adapt to a wide range of censorship mechanisms, substantially enhancing usability and generalizability.

Abstract

The escalating arms race between Internet censorship and evasion has driven censors to evolve from static rule-based filtering to sophisticated deep learning-based traffic analysis. While recent automated evasion tools have attempted to counter this by leveraging stochastic search and programmable heuristics, they continue to suffer from insufficient evasion robustness across diverse censorship modalities and poor usability due to complex, mechanism-specific configurations that require manual fitness tuning or domain-specific languages. In this paper, we propose a paradigm shift that reframes censorship evasion as a semantic image-to-image editing task, allowing users to execute it with a single prompt. We introduce FlowPaint, a novel generative framework that leverages the "world knowledge" of large diffusion models to automatically reshape censored traffic into benign patterns. FlowPaint utilizes an instruction-tuned diffusion architecture to perform semantic editing on network flows. Evaluations against both industrial-grade rule-based middleboxes and learning-based classifiers demonstrate that FlowPaint outperforms existing censorship evasion baselines, enabling users to counter diverse censorship paradigms solely by varying natural language instructions

Problem

• censorship evasion
• traffic analysis
• usability
• robustness
• network flows

Innovation

• generative diffusion models
• censorship evasion
• semantic image-to-image editing
• instruction-tuned diffusion
• FlowPaint