Secure Entity Resolution

Private Set Intersection (PSI) is the foundational cryptographic protocol for secure entity resolution. It allows two or more parties to compute the intersection of their private datasets—such as lists of user IDs—without revealing any information about items not in the intersection. This is achieved using techniques like oblivious transfer, homomorphic encryption, or bloom filters. For example, two hospitals can discover their shared patients for a collaborative study without disclosing their full patient rosters.

Private Information Retrieval (PIR) enables a client to retrieve a specific data item from a server-held database without the server learning which item was retrieved. In the context of entity resolution, PIR can be used after PSI to privately fetch the feature vectors associated with the intersected IDs. Symmetric PIR and single-server PIR are common variants, though they involve significant communication or computational overhead. This mechanism ensures that the act of data alignment itself does not leak query patterns.

Oblivious Transfer (OT) is a protocol where a sender transmits some of its many pieces of information to a receiver, but remains oblivious to which pieces were received. 1-out-of-N OT is frequently used as a building block for more complex PSI protocols. For entity resolution, OT can facilitate the secure matching of records based on encrypted identifiers. Modern, highly efficient OT extension protocols allow millions of these transfers with minimal overhead, making them practical for large-scale VFL workflows.

Homomorphic Encryption (HE) allows computations to be performed directly on encrypted data. For secure entity resolution, Partially Homomorphic Encryption schemes like Paillier or ElGamal can be used to compute set intersections on ciphertexts. A party can encrypt its IDs, another party can perform computations on those ciphertexts using its own set, and only the encrypted result of the match is revealed. While computationally intensive, HE provides a very strong security model, as raw identifiers are never decrypted by other parties.

Bloom Filters are space-efficient probabilistic data structures used for set membership tests. For entity resolution, parties can encode their private sets into Bloom Filters, which are then exchanged and intersected. To handle typos or variations in records (e.g., 'Jon' vs 'John'), fuzzy matching techniques like Locality-Sensitive Hashing (LSH) are combined with cryptographic protocols. This allows for approximate matching of entities, which is crucial for real-world datasets with inconsistent formatting.

Secure Multi-Party Computation (MPC) is a general cryptographic framework that allows multiple parties to jointly compute a function over their private inputs while keeping those inputs concealed. Entity resolution can be formulated as an MPC problem. Garbled Circuits and Secret Sharing are two primary MPC techniques used for this. For example, using additive secret sharing, each party splits its entity identifiers into random shares distributed among others; the intersection is then computed collaboratively on these shares without reconstructing the original values.

Banks, credit card companies, and insurance providers use secure entity resolution to collaboratively identify fraudulent actors without sharing sensitive customer transaction data. By securely aligning customer IDs, institutions can build a composite behavioral profile of a user across multiple services to detect sophisticated, cross-institutional fraud patterns.

• Example: A bank and a telecom provider align user entities to detect if a new credit card application uses a phone number associated with known synthetic identity fraud in the telecom's records.
• Benefit: Dramatically reduces false positives and identifies organized crime rings while maintaining strict compliance with financial privacy regulations like GDPR and GLBA.

Hospitals, research institutes, and pharmaceutical companies employ secure entity resolution to create longitudinal patient records or cohort studies from disparate data sources. This enables cross-institutional clinical studies and the training of diagnostic models via Vertical Federated Learning without transferring Protected Health Information (PHI).

• Example: Aligning patient records from a hospital's EHR system with genomic data from a research lab to study genotype-phenotype correlations for a rare disease.
• Benefit: Accelerates medical research and personalized medicine initiatives while preserving patient confidentiality under HIPAA and other global health data laws.

Brands, publishers, and ad platforms use secure entity resolution to measure campaign effectiveness and understand customer journeys across devices and channels without merging raw user datasets. This solves the identity fragmentation problem in a post-cookie landscape.

• Example: A retailer and a streaming service securely align user entities to measure if viewers of a product's ad later purchased that product online, calculating true Return on Ad Spend (ROAS).
• Benefit: Enables accurate attribution modeling and audience insights while respecting user privacy preferences and regulations like CCPA/CPRA.

Manufacturers, shippers, and retailers apply secure entity resolution to track goods and components across a multi-party supply chain. By aligning entity identifiers like serial numbers or batch IDs, parties can collaboratively predict delays, optimize inventory, and ensure provenance without exposing proprietary operational data.

• Example: An automotive manufacturer, a parts supplier, and a logistics firm align shipment entities to build a federated model that predicts customs clearance delays based on combined—but private—features from each party's records.
• Benefit: Increases supply chain resilience and efficiency through collaborative intelligence while protecting competitive business information.

Government agencies and international bodies use secure entity resolution for threat intelligence sharing and anti-money laundering (AML) efforts. Entities like individuals, organizations, or vessels can be aligned across classified databases from different agencies or allied nations to uncover hidden networks.

• Example: Customs, immigration, and financial intelligence units from multiple countries align entity records to trace the cross-border flow of funds linked to terrorist organizations.
• Benefit: Enhances national security and regulatory enforcement through secure, multi-source data collaboration, operating within strict legal and sovereignty frameworks.

Municipal departments and utility providers leverage secure entity resolution to integrate sensor data from vertically partitioned systems—such as traffic cameras, energy grids, and public transit—to optimize city operations. Entities like geographic zones or asset IDs are aligned to create unified models for predictive maintenance and resource allocation.

• Example: Aligning sensor entities from the water department (pipe pressure) and the transportation department (road vibration) to collaboratively predict and locate underground pipe failures.
• Benefit: Enables holistic, data-driven urban management and the development of digital twins without centralizing control of critical infrastructure data, mitigating single-point security risks.

A cryptographic protocol that allows two or more parties to compute the intersection of their private datasets (e.g., sets of user IDs) without revealing any information about items not in the intersection. It is the primary cryptographic primitive for performing secure entity alignment.

• Purpose: Enables parties in VFL to discover their common users/customers without exposing their full lists.
• Example: A bank and an e-commerce company can find their shared customers to train a joint credit model, without either party learning the other's full customer base.
• Methods: Common protocols include Diffie-Hellman-based PSI, Efficient Circuit-based PSI, and Homomorphic Encryption-based PSI.

A form of encryption that allows computations to be performed directly on ciphertext, generating an encrypted result which, when decrypted, matches the result of the operations as if they had been performed on the plaintext.

• Application in VFL/SER: Used to perform secure computations on aligned data. For example, a feature owner can send encrypted intermediate outputs to a label owner, who can perform model computations on them without decryption.
• Benefit: Provides strong confidentiality guarantees during the training and inference protocols, complementing the alignment provided by PSI.
• Trade-off: Introduces significant computation and communication overhead due to the size and complexity of ciphertext operations.

A rigorous mathematical framework for quantifying and controlling the privacy loss incurred when an individual's data is included in a computation. It guarantees that the output of an algorithm does not reveal much about any single input.

• Relation to SER: While SER/PSI aligns entities, DP can be applied afterward to the training process (e.g., by adding calibrated noise to gradients or aggregated statistics) to provide an additional, post-alignment privacy guarantee.
• Key Parameter: Epsilon (ε), the privacy budget, which bounds the maximum amount of information leakage. A smaller ε implies stronger privacy.
• Use Case: In VFL, DP can be applied by the label owner before updating the global model, protecting against inference attacks based on the final model parameters.

A subfield of cryptography that enables multiple parties to jointly compute a function over their private inputs while keeping those inputs concealed from each other.

• Broader Scope: While PSI is a specific MPC protocol for set intersection, general MPC can be used for the entire VFL training loop, including secure forward/backward propagation and gradient aggregation.
• Mechanism: Parties interact following a protocol that distributes the computation, often using secret sharing techniques, so that no single party sees another's raw data or complete intermediate values.
• Advantage: Can provide end-to-end cryptographic security for the VFL process, from entity alignment through to model training and inference.