Reproduce “AgentStealth: Reinforcing Large Language Model for Anonymizing User-generated Text” on OpenResearch

About this paper

Chenyang Shao, Tianxing Li, Chenhao Pu, Fengli Xu, Yong Li

The AgentStealth framework from Tsinghua University reinforces smaller-scale language models (SLMs) to anonymize user-generated text directly on edge devices, mitigating privacy risks from author profiling without cloud dependency. This approach yields a 12.3% improvement in anonymization performance and a 6.8% enhancement in text utility compared to existing methods.

Key insights

• An adversarial anonymization workflow, enhanced by in-context contrastive learning and adaptive utility prompts, effectively distills knowledge and generates high-quality training data for SLMs.
• Jointly training an SLM in a dual role—as both a privacy defender and an attribute attacker—creates a powerful internal adversary for self-reinforcing learning.
• Leveraging the SLM's own enhanced attack module as the reward signal for reinforcement learning allows for robust and continuous improvement of anonymization capabilities without external attacker models.

Summary by alphaXiv.