Subject-level Inference for Realistic Text Anonymization Evaluation

2026-04-22

AI Summary

Current evaluations of text anonymization rely on fragment-level metrics that fail to capture an adversary’s actual ability to re-identify individuals and overlook privacy risks in multi-subject scenarios. This work proposes SPIA, the first individual-centric privacy evaluation benchmark, which introduces subject-level personally identifiable information (PII) inference metrics based on 675 legal and online documents annotated for multiple subjects. Experimental results reveal that even when over 90% of PII fragments are masked, individuals remain identifiable at a rate as high as 67%, demonstrating that conventional approaches substantially overestimate anonymization effectiveness. Furthermore, anonymization targeting specific subjects can inadvertently heighten exposure risks for non-target subjects, underscoring the need for holistic privacy assessment frameworks.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Current text anonymization evaluation relies on span-based metrics that fail to capture what an adversary could actually infer, and assumes a single data subject, ignoring multi-subject scenarios. To address these limitations, we present SPIA (Subject-level PII Inference Assessment), the first benchmark that shifts the unit of evaluation from text spans to individuals, comprising 675 documents across legal and online domains with novel subject-level protection metrics. Extensive experiments show that even when over 90% of PII spans are masked, subject-level inference protection drops as low as 33%, leaving the majority of personal information recoverable through contextual inference. Furthermore, target-subject-focused anonymization leaves non-target subjects substantially more exposed than the target subject. We show that subject-level inference-based evaluation is essential for ensuring safe text anonymization in real-world settings.

Problem

• text anonymization
• subject-level inference
• PII protection
• multi-subject scenarios
• privacy evaluation

Innovation

• subject-level evaluation
• text anonymization
• PII inference
• contextual inference
• privacy benchmark