A Privacy-Preserving Approach to Extraction of Personal Information through Automatic Annotation and Federated Learning

Rajitha Hathurusinghe, Isar Nejadgholi, Miodrag BolicNAACL 2021

TL;DR

This paper introduces WikiPII, a novel, automatically annotated dataset for Personally Identifiable Information (PII) extraction, designed to address privacy concerns. They demonstrate that even with noisy labels, a BERT-based Named Entity Recognition (NER) model can achieve good performance when trained on a large dataset, and explore federated learning to enable privacy-preserving collaborative training.

What Problem Does This Paper Address?

Extracting Personally Identifiable Information (PII) from text is crucial for many industries, including healthcare and legal services, often for de-identification purposes. However, manually annotating large datasets for PII extraction is time-consuming and expensive. Furthermore, sharing sensitive PII data for annotation and model training raises significant privacy concerns. Existing automated methods often struggle with generalization, necessitating domain-specific datasets. This paper tackles the challenge of creating effective PII extraction models while preserving data privacy and reducing the cost of manual annotation.

Methods

The core of the paper's methodology revolves around two key areas: automatic annotation and federated learning.

  • Automatic Annotation with WikiPII: The authors created a dataset called WikiPII by leveraging Wikipedia biography pages. They extracted PII entities (like birth dates, parent names, spouse names, children names, and education institutes) from the infoboxes of these pages. This infobox data was then used to programmatically annotate the free text of the biography pages. This approach generates a large, albeit noisy, annotated dataset. They acknowledge that this automatic annotation can lead to label noise but argue that with a sufficiently large dataset, machine learning models can compensate for it.
  • Federated Learning for Privacy-Preserving Training: To address the privacy concerns of training PII extraction models, the paper explores federated learning (FL). FL allows models to be trained across multiple decentralized data sources (e.g., different organizations) without the raw data ever leaving its original location. They implemented this using the PySyft framework, which enables remote tensor operations. They investigated two FL scenarios:
  • Federated/Central: A trusted central operator receives data batches from remote holders.
  • Federated/Remote: A mistrusted central operator sends the model to remote data holders for local training, with only model updates being sent back. This scenario is particularly relevant for highly sensitive data.

Datasets

  • WikiPII Dataset: This is a newly introduced dataset comprising 77,703 sentences from Wikipedia biography pages, automatically annotated for 5 classes of personal information: Birth Date (BD), Parents (PR), Spouse(s) (SP), Children (CH), and Education (ED). This dataset is made publicly available for research.
  • CoNLL-2003: This is a standard, widely-used dataset for Named Entity Recognition, used here for comparison and to evaluate the federated learning approach on gold-standard quality annotations.

Key Insights and Ideas

  • Large, Noisy Datasets Can Work: The paper strongly suggests that deep learning models, particularly BERT, are robust to label noise when trained on sufficiently large datasets. This is a critical insight for making PII extraction more scalable and cost-effective.
  • Infoboxes as a Source for Automatic Annotation: Leveraging structured data like Wikipedia infoboxes is an effective and inexpensive way to generate large-scale, automatically annotated datasets for PII extraction, even if the annotations are not perfect.
  • Federated Learning Preserves Privacy Without Sacrificing Performance (Mostly): The study demonstrates that federated learning can be a viable strategy for training PII extraction models without compromising data privacy. While the "federated/remote" scenario (with a mistrusted central operator) showed a slight performance drop, especially with noisy data, it still achieved comparable results to centralized training.
  • Trade-off Between Distributed Training and Data Volume: The authors highlight a trade-off: distributed training might introduce some performance degradation (e.g., due to weight compression), but it enables access to a larger aggregate dataset, which in turn boosts performance.

Key Results

  • Automated Annotation Performance: The rule-based automatic annotation tool for WikiPII showed high noise levels. However, with partial evaluation metrics, it correctly identified about half of the entities, and with strict matching, about one-third were perfectly annotated.
  • BERT Model Performance: A BERT-based NER model trained on the automatically annotated WikiPII dataset achieved an F1-score of 0.73 (strict matching) on the manually annotated test set. This is considered an acceptable performance, especially given the noisy training data.
  • Federated Learning Results:
    • In the "federated/central" scenario, the F1-scores for both CoNLL-2003 and WikiPII were very close to typical centralized training.
  • In the "federated/remote" scenario, the F1-scores were slightly lower, particularly for the noisy WikiPII dataset, but still demonstrated the feasibility of this privacy-preserving approach. For WikiPII, the F1-score was 0.70 ±0.01 compared to 0.70 ±0.006 for centralized training.
  • Effect of Dataset Size: Increasing the dataset size through more participating workers in federated learning generally led to higher F1-scores, reinforcing the idea that more data can overcome noise.

Limitations and Future Work

  • Noise in Automatic Annotation: The paper acknowledges that the automatic annotation process introduces label noise, which can impact model performance. While they show deep learning can handle this with large datasets, it remains a challenge.
  • Performance Drop in "Federated/Remote": The "federated/remote" scenario, while privacy-preserving, showed a performance drop, especially with noisy data. The authors attribute this to potential information loss during model transfer (simplification, serialization, compression).
  • Adversarial Attacks: The authors note that even highly accurate NER models can be vulnerable to adversarial attacks, suggesting that the current system is best envisioned as an assistant to human annotators rather than a fully autonomous solution.
  • Future Work:
    • Rigorous analysis of the effect of object transfers (simplification, serialization, compression) in distributed systems on model precision.
  • Exploring ensembles of different techniques to improve the utility of the PII extraction system.
  • Investigating the impact of imbalanced data distribution in federated learning more deeply.

Summary updated

BibTeX
@inproceedings{hathurusinghe2021privacy,
  author = {Rajitha Hathurusinghe and Isar Nejadgholi and Miodrag Bolic},
  title = {{A Privacy-Preserving Approach to Extraction of Personal Information through Automatic Annotation and Federated Learning}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure