Subject-level Inference for Realistic Text Anonymization Evaluation

TL;DR

This paper introduces SPIA, a new benchmark for evaluating text anonymization at the subject level, revealing significant vulnerabilities even when most PII spans are masked.

Contribution

SPIA shifts evaluation from span-based metrics to subject-level inference, providing novel protection metrics and exposing limitations of current anonymization methods.

Findings

Over 90% of PII spans masked, but subject inference protection drops to 33%.

Contextual inference can recover most personal information.

Target-focused anonymization exposes non-target subjects more.

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.