# DP-Fusion: Token-Level Differentially Private Inference for Large Language Models

**Authors:** Rushil Thareja, Preslav Nakov, Praneeth Vepakomma, Nils Lukas  
**arXiv:** [2507.04531](https://arxiv.org/abs/2507.04531)  
**Date:** 2025-07-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2507.04531v1/

## Key points

- Researchers developed DP-Fusion as a token-level differentially private inference mechanism for large language models (LLMs) to address sensitive information leakage in text generation.
- DP-Fusion offers fine-grained control over the trade-off between privacy and utility by partitioning sensitive tokens into privacy groups and blending output distributions.
- The parameter ε determines the level of privacy protection, with DP-Fusion willing to incur higher computational costs for an improved privacy-utility balance.
- The methodology of DP-Fusion can be extended to other data types such as images and audio, representing a step towards more robust differential privacy mechanisms against real-world threats like LOSS attacks in probabilistic models used for data privacy.

## Summary

, , , , 


In their study "DP-Fusion: Token-Level Differentially Private Inference for Large Language Models," researchers Rushil Thareja, Preslav Nakov, Praneeth Vepakomma, and Nils Lukas address sensitive information leakage from large language models (LLMs) in text generation. They introduce DP-Fusion, a token-level DPI mechanism that balances privacy and utility. By partitioning sensitive tokens into privacy groups and blending output distributions, DP-Fusion offers fine-grained control over the trade-off between privacy and utility. The parameter ε determines the level of privacy protection. One key aspect that sets DP-Fusion apart is its willingness to incur higher computational costs for improved privacy-utility balance. The researchers envision extending this methodology to other data types such as images and audio. DP-Fusion represents a step towards more robust differential privacy mechanisms against real-world threats like LOSS attacks in probabilistic models used for data privacy. A token-level DPI mechanism for balancing privacy and utility in large language models. The issue addressed by researchers in their study on DP-Fusion. Techniques used to protect sensitive information in various data types. The goal of DP-Fusion, achieved through partitioning sensitive tokens into groups and blending output distributions. A method for protecting sensitive information while maintaining statistical accuracy in data analysis.