GRASP: Reinforcing Language Model Anonymizers with Group Relative Policy Optimization
Summary:The paper introduces GRASP (Group-Relative Anonymization via Self-refinement Policy-optimization), a novel approach for reinforcing on-device language model anonymizers. GRASP employs a single small model as both the anonymizer, adversary, and utility judge, trained against a self-generated reward to hide sensitive attributes while preserving meaning. Compared to methods based on supervised fine-tuning and Direct Preference Optimization (DPO), GRASP achieves a better privacy-utility trade-off an
Background and Challenges
Large Language Models (LLMs) can infer sensitive personal attributes such as age, location, and occupation from ordinary text, posing a privacy risk to everyday writing. Existing adversarial anonymization methods use powerful language models to rewrite text to hide sensitive information, but they require strong models and send private text to third-party servers during inference, which contradicts the purpose of anonymization.
Overview of GRASP
GRASP (Group-Relative Anonymization via Self-refinement Policy-optimization) reinforces on-device anonymizers using Group Relative Policy Optimization. Its key innovations include:
- Single-Model Multi-Role Design: A single small model acts as both the anonymizer, adversary, and utility judge.
- Self-Generated Reward Mechanism: The model is trained against a self-generated reward that hides sensitive attributes while preserving meaning.
- Reward Hacking Prevention: The design includes mechanisms to prevent the model from exploiting shortcuts or cheating to obtain rewards.
Experimental Results and Advantages
Trained on Llama-3.1-8B, GRASP significantly improves the privacy-utility trade-off compared to the DPO-distilled baseline. In comparison with adversarial anonymization methods driven by frontier models like Gemini 2.5 Flash and Claude, GRASP achieves a comparable or better overall trade-off while removing substantially more private information. Additionally, GRASP operates entirely on-device at roughly 1% of the GPT-4o teacher's cost.
Technical Highlights
- Innovative Policy Optimization Method: GRASP employs a novel group relative policy optimization approach for more efficient anonymization.
- Single-Model Multi-Role Design: Simplifies the model architecture and reduces computational costs.
- On-Device Operation: Eliminates the need for third-party servers, enhancing privacy protection.
- Self-Generated Reward Mechanism: Improves the model's performance in the privacy-utility trade-off.
Industry Impact and Developer Recommendations
GRASP provides a new technical direction for the privacy protection field, particularly for applications that handle sensitive data, such as healthcare, finance, and social networks. For developers, GRASP offers an efficient, low-cost anonymization solution that helps protect user privacy while maintaining text usability.
Future Outlook
In the future, GRASP is expected to be applied in more fields, such as intelligent assistants, chatbots, and content moderation systems. As the technology continues to evolve, GRASP's performance and efficiency will further improve, providing stronger support for privacy protection.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-08-10)
Tags: #Privacy Protection #Large Language Models #On-Device Computing #Policy Optimization
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments