Dchi-Stencil: A Differential Privacy Mechanism for Interacting with LLMs

ICLR 2026 Conference Submission7703 Authors

Keywords: LLMs, Differential, Privacy, Inference

Abstract:

The use of language models as remote services requires transmitting private information to external providers, raising significant privacy concerns.
This process not only risks exposing sensitive data to untrusted service providers but also leaves it vulnerable to interception by eavesdroppers.
Existing privacy-preserving methods for natural language processing (NLP) interactions primarily rely on semantic similarity, overlooking the role of contextual information.
In this work, we introduce $d_\chi$-Stencil, a novel token-level privacy-preserving mechanism that integrates contextual and semantic information while ensuring strong privacy guarantees under the $d_\chi$ differential privacy framework, achieving $2\epsilon$-$d_\chi$-privacy.
By incorporating both semantic and contextual nuances,$d_\chi$-Stencil achieves a robust balance between privacy and utility.
We evaluate $d_\chi$-Stencil using state-of-the-art language models and diverse datasets, achieving comparable and even better trade-off between utility and privacy compared to existing methods.
This work highlights the potential of $d_\chi$-Stencil to set a new standard for privacy-preserving NLP in modern, high-risk applications.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 7703