De-Identification of Sensitive Personal Data in Datasets Derived from IIT-CDIP

Stefan Larson, Nicole Lima, Santiago Diaz, Amogh Manoj Joshi, Siddharth Betala, Jamiu Suleiman, Yash Mathur, Kaushal Prajapati, Ramla Alakraa, Junjie Shen, Temi Okotore, Kevin LeachEMNLP 2024 · DOI

Introduction

In the field of Document AI, datasets derived from the IIT-CDIP collection—such as RVL-CDIP, DocVQA, and FUNSD—serve as the foundational benchmarks for training models to classify, read, and understand complex document images. However, these datasets, sourced from real-world legal settlements in the tobacco industry, contain a significant and previously unquantified amount of unredacted Personally Identifiable Information (PII).

As Large Language Models (LLMs) and diffusion models demonstrate an increasing tendency to "memorize" and potentially leak training data, the presence of Social Security Numbers (SSNs), home addresses, and birth dates in public benchmarks poses severe ethical and security risks. This research quantifies the extent of this leakage and proposes a "pseudonymization" pipeline to scrub sensitive data while preserving the utility of the datasets for machine learning research.

Quantifying the Risk: Manual Audit and Tool Evaluation

To understand the scope of the problem, researchers conducted a manual audit across five major datasets. The audit revealed a startling 16,396 instances of sensitive PII, including 2,428 US Social Security Numbers, over 13,000 birth dates, and more than 4,000 home addresses.

Overview of PII detection and the pseudonymization process.

An evaluation of "off-the-shelf" PII detectors (including Microsoft Presidio, Amazon Comprehend, and Google DLP) found that automated tools struggle with the noisy, low-context environment of scanned documents. While some tools caught a high percentage of SSNs, they failed to detect up to 30% of other PII types, and most tools were unable to identify 7 out of the 11 PII categories tracked in the study.

Methodology: The Pseudonymization Pipeline

To address these leaks, the researchers developed a de-identification pipeline that goes beyond simple redaction. They compared three strategies:

  1. Black/White Redaction: Covering PII with solid color boxes.
  2. Pseudonymization: Replacing real PII with synthetic, plausible data (e.g., fake names and SSNs generated via the Faker library).
  3. Realistic Rendering: Using the Augraphy library to apply noise, jitter, and low-ink effects to synthetic text, ensuring it matches the visual quality of the original scanned document.

By replacing a real SSN with a fake one that matches the original document's style, the researchers "hide" the sensitive data among realistic decoys, maintaining the document's visual structure.

Impact on Data Utility and Model Performance

A critical concern in data de-identification is whether the changes render the dataset useless for training. The study found that pseudonymization is significantly better at preserving data utility than traditional black-box redaction. Documents with synthetic replacements were 61.6% more similar to their originals in CLIP embedding space than those with white-box redactions.

Distribution of redacted area relative to total document area.

Furthermore, the impact on model performance is minimal. Because PII typically occupies less than 2% of a document's total area (as shown in the distribution above), redaction does not significantly degrade the accuracy of document classifiers like the Document Image Transformer (DiT). While black redactions caused a slight drop in model confidence, the pseudonymized versions allowed models to maintain high-confidence classifications.

Conclusion

This work highlights a critical oversight in the curation of popular Document AI benchmarks. By implementing a pseudonymization pipeline that combines synthetic data generation with realistic rendering, researchers can protect individual privacy without compromising the scientific value of the data. The team is working to release these redacted versions to establish a new, safer standard for the Document AI community.

Summary updated

BibTeX
@inproceedings{larson2024de,
  author = {Stefan Larson and Nicole Lima and Santiago Diaz and Amogh Manoj Joshi and Siddharth Betala and Jamiu Suleiman and Yash Mathur and Kaushal Prajapati and Ramla Alakraa and Junjie Shen and Temi Okotore and Kevin Leach},
  title = {{De-Identification of Sensitive Personal Data in Datasets Derived from IIT-CDIP}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.emnlp-main.1198},
}
Figure