You Only Anonymize What Is Not Intent-Relevant: Suppressing Non-Intent Privacy Evidence

TL;DR

This work tackles inference-based privacy threats in text by moving beyond uniform anonymization to intent-aware exposure control. It introduces IntentAnony, an end-to-end framework comprising pragmatic intent recognition, privacy inference evidence chains, and scene–intent level governance to allocate per-attribute exposure budgets, enabling selective suppression of non-intent privacy cues while preserving intent-relevant content. Across two Reddit-style datasets and multiple backbone LLMs, IntentAnony demonstrates a stronger privacy-utility balance than baselines, with better semantic preservation and minimal intent drift. The findings highlight the practical potential of intent-aware anonymization for real-world language use, while acknowledging limitations and directions for future work toward uncertainty-aware intent modeling and formal privacy guarantees.

Abstract

Anonymizing sensitive information in user text is essential for privacy, yet existing methods often apply uniform treatment across attributes, which can conflict with communicative intent and obscure necessary information. This is particularly problematic when personal attributes are integral to expressive or pragmatic goals. The central challenge lies in determining which attributes to protect, and to what extent, while preserving semantic and pragmatic functions. We propose IntentAnony, a utility-preserving anonymization approach that performs intent-conditioned exposure control. IntentAnony models pragmatic intent and constructs privacy inference evidence chains to capture how distributed cues support attribute inference. Conditioned on intent, it assigns each attribute an exposure budget and selectively suppresses non-intent inference pathways while preserving intent-relevant content, semantic structure, affective nuance, and interactional function. We evaluate IntentAnony using privacy inference success rates, text utility metrics, and human evaluation. The results show an approximately 30% improvement in the overall privacy--utility trade-off, with notably stronger usability of anonymized text compared to prior state-of-the-art methods. Our code is available at https://github.com/Nevaeh7/IntentAnony.

Figures (8)

• Figure 1: Anonymization under inference-based privacy threats. Compared with uniform rewriting methods that obscure sensitive details at the cost of communicative intent, IntentAnony conditions anonymization on pragmatic intent and selectively suppresses non-intent privacy evidence, preserving semantic meaning and interactional function while reducing inference risk.
• Figure 2: Overview of the proposed intent-aware anonymization framework. The pipeline includes four stages: (1) pragmatic intent recognition to identify communicative intents in the input text; (2) privacy inference evidence chain construction that organizes sensitive attributes into intent-grounded evidence chains; (3) scene--intent level privacy exposure governance for determining appropriate anonymization levels; and (4) evidence chain anonymization, which selectively rewrites or removes non-intent evidence while preserving intent-relevant content.
• Figure 3: Privacy--utility trade-off across five privacy granularity levels (L0 to BAN), showing the relationship between inference attack accuracy and mean text utility on multiple commercial language models.
• Figure 4: Distribution of semantic similarity scores between anonymized and original texts on the PersonalReddit dataset. IntentAnony yields distributions more concentrated toward higher values than baseline methods, indicating more consistent preservation of original semantics and communicative intent.
• Figure 5: Comparison of intent preservation across different anonymization methods, measured by Intent Overlap and Stability F1, where higher scores indicate better alignment with the original communicative intent.