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.
Cite
@article{arxiv.2604.21211,
title = {Subject-level Inference for Realistic Text Anonymization Evaluation},
author = {Myeong Seok Oh and Dong-Yun Kim and Hanseok Oh and Chaean Kang and Joeun Kang and Xiaonan Wang and Hyunjung Park and Young Cheol Jung and Hansaem Kim},
journal= {arXiv preprint arXiv:2604.21211},
year = {2026}
}