Local Language Models for Context-Aware Adaptive Anonymization of Sensitive Text

TL;DR

This paper presents a context-aware, adaptive anonymization framework using local language models to improve privacy in qualitative research data, outperforming manual and rule-based methods.

Contribution

It introduces the Structured Framework for Adaptive Anonymizer (SFAA) utilizing local LLMs for sensitive data detection and anonymization, enhancing accuracy and consistency.

Findings

Phi model detected over 91% of sensitive data

Phi maintained 94.8% sentiment consistency

LLMs identified more sensitive data than human reviewers

Abstract

Qualitative research often contains personal, contextual, and organizational details that pose privacy risks if not handled appropriately. Manual anonymization is time-consuming, inconsistent, and frequently omits critical identifiers. Existing automated tools tend to rely on pattern matching or fixed rules, which fail to capture context and may alter the meaning of the data. This study uses local LLMs to build a reliable, repeatable, and context-aware anonymization process for detecting and anonymizing sensitive data in qualitative transcripts. We introduce a Structured Framework for Adaptive Anonymizer (SFAA) that includes three steps: detection, classification, and adaptive anonymization. The SFAA incorporates four anonymization strategies: rule-based substitution, context-aware rewriting, generalization, and suppression. These strategies are applied based on the identifier type and the risk level. The identifiers handled by the SFAA are guided by major international privacy and research ethics standards, including the GDPR, HIPAA, and OECD guidelines. This study followed a dual-method evaluation that combined manual and LLM-assisted processing. Two case studies were used to support the evaluation. The first includes 82 face-to-face interviews on gamification in organizations. The second involves 93 machine-led interviews using an AI-powered interviewer to test LLM awareness and workplace privacy. Two local models, LLaMA and Phi were used to evaluate the performance of the proposed framework. The results indicate that the LLMs found more sensitive data than a human reviewer. Phi outperformed LLaMA in finding sensitive data, but made slightly more errors. Phi was able to find over 91% of the sensitive data and 94.8% kept the same sentiment as the original text, which means it was very accurate, hence, it does not affect the analysis of the qualitative data.