Protecting Private Information While Preserving Semantic Integrity in LLM-Assisted Systems: It can be done.

ACL ARR 2025 February Submission5298 Authors

Abstract: With the increasing use of AI-assisted systems, there is growing concern over privacy leaks, especially when users share sensitive personal data in interactions with Large Language Models (LLMs). Conversations shared with these models may contain Personally Identifiable Information (PII) that could be exposed. To address this issue, we present the LOPSIDED framework, a semantically-aware privacy agent designed specifically for remote LLMs. Our approach involves pseudonymizing requests during inference and de-pseudonymizing them once the response is generated, ensuring that sensitive information is protected without compromising the quality of the LLM's output. We evaluate our approach using real-world conversations sourced from ShareGPT. Furthermore, we augment and annotate this data to determine whether named entities are relevant to the prompt and impact the LLM's output. Our analysis reveals that our method reduces utility errors by a factor of 5 compared to baseline techniques, all while maintaining privacy.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: security/privacy, evaluation methodology, named entity recognition and relation extraction

Contribution Types: Model analysis & interpretability, Approaches to low-resource settings, Publicly available software and/or pre-trained models, Data resources

Languages Studied: English

Submission Number: 5298