Broad and Efficient PII Detection with Unified Labels

Key points

• Personally Identifiable Information (PII) detection models must efficiently find identifiers across diverse contexts while maintaining a low false-positive rate (FPR) on high-volume traffic.1
• Recent detectors increasingly rely on complex architectures that raise deployment cost, yet many are fine-tuned on only a few synthetic datasets that limit their coverage.1
• Combining public datasets is challenging because their taxonomies differ, and even after categories are mapped to a common schema, labels can disagree on what counts as a detection, producing inconsistent training and evaluation signals.1
• Evaluated on the AI4Privacy and Nemotron-PII benchmarks at a low-FPR operating point, our model achieves the highest recall against larger baselines in 11 of 14 comparisons under our labels and 9 of 14 under the original labels.1

Sources (1)

• [1]Broad and Efficient PII Detection with Unified Labels
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 07:47 PM

  To address these problems, we present a practical 141M-parameter encoder-based detector trained on 1.12 million examples from 77 public datasets, relabeled under a detailed definition of broad contextual detection for 12 structured PII categories.

  Personally Identifiable Information (PII) detection models must efficiently find identifiers across diverse contexts while maintaining a low false-positive rate (FPR) on high-volume traffic.

Extractive summary: sentences quoted from the sources.