Researchers unveil GRASP for private AI text anonymization
GRASP hides personal details in text while keeping meaning intact—using 1% of the cost of GPT-4o.
Researchers Sajjad Ghiasvand and Nader Sehatbakhsh introduce GRASP, a novel approach to anonymizing text generated by large language models (LLMs). Traditional methods either require sending sensitive data to third-party models or fail to balance privacy and utility effectively. GRASP addresses this by using a single small model that acts as an anonymizer, adversary, and utility judge, trained with Group Relative Policy Optimization (GRPO). This method reinforces the model online, optimizing for both privacy and meaning preservation directly, rather than relying on offline imitation learning like DPO.
GRASP was trained on the Llama-3.1-8B model and evaluated against state-of-the-art adversarial anonymizers, including those powered by Gemini 2.5 Flash and Claude 3.5. Results show GRASP consistently improves the privacy-utility trade-off, removing significantly more private information while running entirely on-device at roughly 1% of the cost of GPT-4o. This breakthrough enables privacy-preserving text generation without external dependencies, making it ideal for sensitive applications.
- GRASP uses Group Relative Policy Optimization (GRPO) to train a single small model for anonymization, adversarial testing, and utility judgment, improving privacy-utility trade-offs.
- Trained on Llama-3.1-8B, GRASP removes more private information than DPO-distilled baselines and rivals frontier models like Gemini 2.5 Flash and Claude 3.5.
- Runs entirely on-device at ~1% of GPT-4o’s cost, eliminating third-party privacy risks.
Why It Matters
GRASP enables secure, cost-effective anonymization of LLM-generated text, protecting privacy without sacrificing performance or control.