Towards Privacy-Preserving LLM Inference via Collaborative Obfuscation (Technical Report)

Researchers have developed AloePri, the first privacy-preserving LLM inference method designed for industrial applications. The system uses collaborative obfuscation to protect input/output data while maintaining 96.5-100% accuracy and resisting state-of-the-art attacks, successfully tested on a 671B parameter model.

• →AloePri is the first privacy-preserving LLM inference method that meets all three industrial requirements: minimal accuracy loss, scalability on heterogeneous hardware, and infrastructure compatibility.
• →The system achieved 0.0-3.5% accuracy loss while maintaining efficiency equivalent to plaintext inference on the 671B parameter Deepseek-V3.1-Terminus model.
• →AloePri successfully resisted state-of-the-art attacks with less than 5% of tokens being recovered during testing.
• →The method uses covariant obfuscation to jointly transform both data and model parameters for enhanced privacy protection.
• →This represents the first practical privacy-preserving solution for large-scale LLM deployments in real-world industrial systems.