GRASP: Reinforcing Language Model Anonymizers with Group Relative Policy Optimization
Summary:GRASP introduces a novel approach to anonymizing text by leveraging Group-Relative Policy Optimization to enhance on-device anonymizers. This technique employs a single small model to act as both the anonymizer and the adversary, optimizing for privacy-utility trade-offs through self-generated rewards. Trained on Llama-3.1-8B, GRASP demonstrates superior performance in preserving meaning while removing sensitive attributes compared to DPO-distilled baselines. It achieves comparable or better res
Background and Challenges
Large Language Models (LLMs) can infer sensitive personal attributes from ordinary text, posing a significant privacy risk. Traditional adversarial anonymization methods rely on powerful language models to rewrite text and hide sensitive information. However, these methods require substantial computational resources and necessitate sending private data to third parties, which undermines the purpose of anonymization.
Core Innovations of GRASP
- Group-Relative Policy Optimization: GRASP enhances on-device anonymizers using Group-Relative Policy Optimization, enabling the model to hide sensitive attributes while preserving the meaning of the text.
- Single Model, Multiple Roles: The technique employs a single small model to act as both the anonymizer and the adversary, optimizing for privacy-utility trade-offs through self-generated rewards.
- On-Device Execution: GRASP operates entirely on-device, significantly reducing the risk of privacy leakage and the need for extensive computational resources.
Experimental Results and Advantages
- Privacy-Utility Trade-off: Trained on Llama-3.1-8B, GRASP outperforms DPO-distilled baselines in terms of privacy-utility trade-offs.
- Removal of Private Information: Compared to state-of-the-art anonymization driven by models like Gemini 2.5 Flash and Claude, GRASP removes substantially more private information while maintaining semantic integrity.
- On-Device Efficiency: The on-device execution capability of GRASP makes it highly advantageous in resource-constrained environments.
Industry Impact and Developer Recommendations
GRASP offers a new approach to privacy protection, particularly in sensitive applications such as healthcare, finance, and social media. It provides developers with an efficient and cost-effective anonymization solution, reducing reliance on powerful computational resources and third-party services.
Future Directions
As privacy concerns continue to grow, GRASP is poised to find applications in a wider range of domains. Future research could further optimize its performance and explore its potential in multilingual environments.
Conclusion
GRASP demonstrates the potential of policy optimization in reinforcing on-device anonymizers, paving the way for new directions in AI-driven privacy protection technologies.
— END —Tags: #Privacy Protection #Large Language Models #On-Device AI #Policy Optimization #Anonymization
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