A privacy-preserving distributed filtering framework for NLP artifacts

Md. Nazmus Sadat, Md Momin Al Aziz, Noman Mohammed, Serguei Pakhomov, Hongfang Liu, Xiaoqian JiangBMC Medical Informatics Decis. Mak 2019 · OpenReview

TL;DR

This paper presents a novel framework for de-identifying clinical notes from multiple distributed sources, addressing the challenge of sharing sensitive medical data for research. By combining private set intersection and homomorphic encryption, it enables collaborative filtering of low-frequency, potentially identifying terms without revealing raw data, thus preserving both privacy and data utility.

What Problem Does This Paper Address?

Sharing medical data is crucial for advancing research and improving healthcare, but privacy concerns, particularly with unstructured clinical notes, pose a significant barrier. While de-identification methods exist, accurately and automatically removing all sensitive information is challenging. Existing approaches often struggle with distributed data sources and maintaining data utility after filtering. This paper tackles the problem of de-identifying clinical notes across multiple institutions in a privacy-preserving manner, ensuring that sensitive information is not exposed during the process.

Methods

The core of this work is a distributed filtering framework that extends a previous single-source method. Instead of directly sharing clinical notes, the proposed approach focuses on identifying and filtering sentences containing potentially identifying terms. This is achieved through a novel secure protocol that combines:

  • Private Set Intersection (PSI): This allows multiple parties to find common elements (in this case, bigrams) in their datasets without revealing their individual datasets to each other.
  • Secure Thresholding: After identifying common bigrams, the system uses homomorphic encryption to securely aggregate the counts of these bigrams across all participating data owners. A predetermined threshold is then applied to identify "globally infrequent" bigrams, which are considered privacy-sensitive.
  • Homomorphic Encryption (FV Scheme): This cryptographic technique allows computations (like addition) to be performed on encrypted data without decrypting it. This is crucial for aggregating bigram counts securely.
  • Ciphertext Packing (SIMD): To mitigate the computational overhead of homomorphic encryption, the paper utilizes ciphertext packing, enabling parallel processing of multiple data points within a single ciphertext, significantly boosting performance.
  • SHA-256 Hashing: Used for initial bigram matching and intersection, with salting to enhance security against attacks.

Datasets

The primary dataset used for evaluation is the MIMIC-III (Medical Information Mart for Intensive Care) database. Specifically, the NOTEEVENTS table was utilized, which contains de-identified clinical notes from approximately 40,000 critical care patients. This dataset is a standard benchmark for research in this area. No new datasets were introduced.

Key Insights and Ideas

  • Distributed De-identification is Feasible: The paper demonstrates that it's possible to de-identify clinical notes collaboratively across multiple institutions without centralizing or exposing raw sensitive data.
  • Homomorphic Encryption for Secure Aggregation: Leveraging homomorphic encryption for secure count aggregation is a key enabler, allowing for privacy-preserving thresholding of bigram frequencies.
  • Ciphertext Packing for Performance: The use of ciphertext packing is a critical optimization that makes the homomorphic encryption approach practical for large-scale datasets.
  • Bigram Frequency as a Privacy Indicator: The core idea of filtering sentences based on the global rarity of bigrams is a clever way to balance privacy and data utility. Rarer bigrams are more likely to be unique identifiers.
  • Generalizability: The framework is designed to be generalizable beyond bigrams to other NLP artifacts like unigrams and n-grams.

Key Results

  • Scalability: Experimental results show that the proposed framework scales well with increasing cardinality of the intersection of sets and the number of data owners. For instance, total processing time increased from 75.38 seconds to 104.11 seconds as the intersection cardinality grew from ~10% to ~14% of the dataset.
  • Data Utility Preservation: The evaluation of concept distribution (using Type Unique Identities - TUIs) showed that the filtered corpus retains enough information for data analysis, with minimal differences compared to the original data.
  • Efficiency: The use of ciphertext packing led to significant time cost reductions compared to naive homomorphic encryption.
  • Security: The framework is designed to be secure against common cryptographic attacks, with a claimed 142-bit security level based on root-Hermite factor analysis for encryption and the use of SHA-256 with salting for hashing.

Limitations and Future Work

  • Computational Cost: While optimized with ciphertext packing, homomorphic encryption still incurs significant computational overhead, especially for large datasets. The paper notes that runtime is dependent on the intersection size and number of data owners.
  • Oversampling for Distribution Maintenance: The paper acknowledges that filtering can alter data distribution and suggests oversampling filtered corpus sentences containing specific TUIs as a potential solution, though this is noted as a standard combinatorial optimization problem not explored in detail.
  • Future Work: The authors suggest that the framework is generalizable to other NLP artifacts and can be further optimized. The open-sourced code and manual are provided for further development and application.

Summary updated

Abstract

Medical data sharing is a big challenge in biomedicine, which often hinders collaborative research. Due to privacy concerns, clinical notes cannot be directly shared. A lot of efforts have been dedicated to de-identifying clinical notes but it is still very challenging to accurately locate and scrub all sensitive elements from notes in an automatic manner. An alternative approach is to remove sentences that might contain sensitive terms related to personal information. A previous study introduced a frequency-based filtering approach that removes sentences containing low frequency bigrams to improve the privacy protection without significantly decreasing the utility. Our work extends this method to consider clinical notes from distributed sources with security and privacy considerations. We developed a novel secure protocol based on private set intersection and secure thresholding to identify uncommon and low-frequency terms, which can be used to guide sentence filtering. As the computational cost of our proposed framework mostly depends on the cardinality of the intersection of the sets and the number of data owners, we evaluated the framework in terms of these two factors. Experimental results demonstrate that our proposed method is scalable in various experimental settings. In addition, we evaluated our framework in terms of data utility. This evaluation shows that the proposed method is able to retain enough information for data analysis. This work demonstrates the feasibility of using homomorphic encryption to develop a secure and efficient multi-party protocol.
BibTeX
@inproceedings{sadat2019privacy,
  author = {Md. Nazmus Sadat and Md Momin Al Aziz and Noman Mohammed and Serguei Pakhomov and Hongfang Liu and Xiaoqian Jiang},
  title = {{A privacy-preserving distributed filtering framework for NLP artifacts}},
  booktitle = {BMC Medical Informatics Decis. Mak},
  year = {2019},
  url = {https://openreview.net/forum?id=c67AoYx6wo},
}
Figure