Researchers have introduced SPIA, a new benchmark for evaluating text anonymization that focuses on individual-level inference rather than just masked text spans. Current methods, even those masking over 90% of personally identifiable information (PII), can still leave significant personal details recoverable through contextual inference. The study also found that anonymizing for a specific target subject can inadvertently expose non-target subjects more severely, highlighting the need for subject-level inference evaluation for real-world safety. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT New evaluation benchmark highlights critical gaps in current text anonymization techniques, potentially impacting data privacy practices in AI.
RANK_REASON Introduces a new benchmark and evaluation methodology for text anonymization.
