Reproduce “Unmasking the Reality of PII Masking Models: Performance Gaps and the Call for Accountability” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Devansh Singh, Sundaraparipurnan Narayanan
Researchers at AI Tech Ethics evaluated popular open-source PII masking models against a new, diverse semi-synthetic dataset, revealing substantial rates of PII non-identification and misclassification. This highlights critical privacy risks and underscores the need for greater accountability and transparency in AI model development and deployment.
Key insights
- Popular open-source PII masking models exhibit significant performance gaps, with collective non-identification of PII in 28% of predictions and incorrect classification for 67% of detected PII.
- Models struggle significantly with contextual entity disambiguation and implicit PII recognition, especially when PII terms have multiple interpretations or lack explicit type references.
- Misclassification rates are high across models and feature dimensions, often stemming from insufficient variation and diversity in the training data rather than merely the absence of explicit contextual cues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.