Privacy-preserving Fuzzy Name Matching for Sharing Financial Intelligence

Harsh Kasyap, Ugur Ilker Atmaca, Carsten Maple, Graham Cormode, Jiancong He

Abstract

Financial institutions rely on data for many operations, including a need to drive efficiency, enhance services and prevent financial crime. Data sharing across an organisation or between institutions can facilitate rapid, evidence-based decision-making, including identifying money laundering and fraud. However, modern data privacy regulations impose restrictions on data sharing. For this reason, privacy-enhancing technologies are being increasingly employed to allow organisations to derive shared intelligence while ensuring regulatory compliance. This paper examines the case in which regulatory restrictions mean a party cannot share data on accounts of interest with another (internal or external) party to determine individuals that hold accounts in both datasets. The names of account holders may be recorded differently in each dataset. We introduce a novel privacy-preserving scheme for fuzzy name matching across institutions, employing fully homomorphic encryption over MinHash signatures. The efficiency of the proposed scheme is enhanced using a clustering mechanism. Our scheme ensures privacy by only revealing the possibility of a potential match to the querying party. The practicality and effectiveness are evaluated using different datasets, and compared against state-of-the-art schemes. It takes around 100 and 1000 seconds to search 1000 names from 10k and 100k names, respectively, meeting the requirements of financial institutions. Furthermore, it exhibits significant performance improvement in reducing communication overhead by 30-300 times.

Introduction

Financial crime, including money laundering and fraud, represents a multi-trillion-dollar global threat. To combat these activities effectively, financial institutions must share "financial intelligence"—specifically, checking if a suspicious individual in one bank's database also holds accounts at another. However, two major hurdles impede this collaboration. First, stringent privacy regulations like GDPR prevent the direct sharing of customer lists. Second, data is often "fuzzy"; names are rarely recorded identically due to spelling variations (e.g., "Jon" vs. "John"), nicknames, or transcription errors, rendering exact matching techniques like standard Private Set Intersection (PSI) insufficient.

Existing solutions for Private Record Linkage often fail to scale to the massive datasets typical of the banking sector or require prohibitive communication overhead. This paper presents a novel system that allows institutions to cross-reference lists of suspicious individuals with high precision without ever revealing the underlying names.

System Architecture and Methods

The proposed scheme utilizes a multi-layered approach to ensure both security and scalability. The process begins with data preparation, where names are broken down into "shingles" (n-grams) and converted into MinHash signatures. This technique compresses the data while preserving Jaccard or Cosine similarity, ensuring that similar names result in similar signatures.

Figure 1: High-level architecture of the privacy-preserving fuzzy name matching system, illustrating the interaction between the Querier (Organisation A) and the Responder (Organisation B).

To perform comparisons without decryption, the system employs CKKS Fully Homomorphic Encryption (FHE). This allows the Responder (Organisation B) to calculate the similarity between an encrypted query from the Querier (Organisation A) and its own local data. The Querier receives only the encrypted results, which they then decrypt to determine if a match exists based on a predefined threshold.

Scaling via Clustering

A primary innovation of this work is the clustering-based scaling mechanism. Comparing a query to every record in a million-name database in the encrypted domain is computationally expensive. To solve this, the Responder pre-clusters their data using K-Means. The Querier sends an encrypted "one-hot" indicator vector to select the most relevant cluster centroid. This narrows the search space significantly while ensuring the Responder remains unaware of which specific cluster was queried.

Evaluation and Key Results

The effectiveness of the scheme was evaluated using three datasets: the North Carolina Voter Registration (NCVR) dataset (1 million records) for scalability, library catalogue datasets for long-string matching, and US Census data for sensitivity analysis against controlled Levenshtein distances.

Performance and Efficiency

  • Speed: The system can search 1,000 names against a 100,000-name database in approximately 1,000 seconds. This performance is suitable for the batch processing workflows common in financial institutions.
  • Communication Overhead: The clustering mechanism achieves a 30–300x reduction in communication costs. For 100 queries against a million names, the data transfer was reduced from 314 GB to approximately 1–2 GB.
  • Precision and Recall: The system prioritizes precision (99.5%+) to avoid the reputational risk of false positives. While recall naturally decreases as names become more distorted (e.g., a Levenshtein distance of 5), it remains near 100% for minor variations (distance of 0 or 1).

Discussion and Future Work

While the system offers a significant leap in efficiency, the use of clustering introduces a "recall gap" where names sitting on cluster boundaries might be missed. Additionally, the Responder requires substantial memory (e.g., 77 GB for 1 million records) to handle large-scale column-wise operations.

Future research will focus on minimizing the recall drop caused by clustering, potentially through the implementation of overlapping clusters or more advanced indexing techniques that remain compatible with FHE constraints. Despite these challenges, the proposed scheme provides a practical, privacy-compliant path for financial institutions to collaborate in the fight against global financial crime.

Summary updated

Abstract

Financial institutions rely on data for many operations, including a need to drive efficiency, enhance services and prevent financial crime. Data sharing across an organisation or between institutions can facilitate rapid, evidence-based decision-making, including identifying money laundering and fraud. However, modern data privacy regulations impose restrictions on data sharing. For this reason, privacy-enhancing technologies are being increasingly employed to allow organisations to derive shared intelligence while ensuring regulatory compliance. This paper examines the case in which regulatory restrictions mean a party cannot share data on accounts of interest with another (internal or external) party to determine individuals that hold accounts in both datasets. The names of account holders may be recorded differently in each dataset. We introduce a novel privacy-preserving scheme for fuzzy name matching across institutions, employing fully homomorphic encryption over MinHash signatures. The efficiency of the proposed scheme is enhanced using a clustering mechanism. Our scheme ensures privacy by only revealing the possibility of a potential match to the querying party. The practicality and effectiveness are evaluated using different datasets, and compared against state-of-the-art schemes. It takes around 100 and 1000 seconds to search 1000 names from 10k and 100k names, respectively, meeting the requirements of financial institutions. Furthermore, it exhibits significant performance improvement in reducing communication overhead by 30-300 times.
BibTeX
@misc{kasyap2024privacy,
  author = {Harsh Kasyap and Ugur Ilker Atmaca and Carsten Maple and Graham Cormode and Jiancong He},
  title = {{Privacy-preserving Fuzzy Name Matching for Sharing Financial Intelligence}},
  year = {2024},
  eprint = {2407.19979},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.19979},
}
Figure