Privacy-Preserving Entity Alignment Without Disclosure: Unlocking Secure Collaboration
Artificial intelligence is entering a phase where collaboration is no longer optional, but essential. The most valuable insights today emerge not from isolated datasets, but from the combination of multiple, distributed sources across organizations, industries, and borders. Yet this opportunity comes with a fundamental constraint: data cannot move.
Privacy regulations, competitive boundaries, and security risks make traditional data centralization not only impractical, but often impossible. In this context, the paper “Sherpa.ai Privacy-Preserving Multi-Party Entity Alignment without Intersection Disclosure for Noisy Identifiers” addresses one of the most critical and least solved challenges in modern AI: how to align and combine data across organizations without ever exposing the data itself.
The Hidden Bottleneck of Collaborative AI
In Vertical Federated Learning, different organizations hold complementary pieces of information about the same entities. A bank may know financial behavior, an insurer risk profiles, and a hospital clinical data. To train a shared model, these datasets must be aligned so that each row corresponds to the same real-world entity.
Figure 1: Illustrative example of entity alignment in VFL: based on the ID, Parties A and B perform private matching and produce Aligned Data without disclosing intersection membership or sensitive values.
Until now, this alignment step has been a weak point.
Traditional approaches such as Private Set Intersection (PSI) reveal which entities are common between both datasets.
Figure 2: Illustration of the PSI protocol. Only the common identifiers (IDs) between the two parties are used to form the shared training dataset.
This seemingly small leak can have severe consequences. Knowing that a customer appears in both a bank and a cancer clinic dataset, for example, may implicitly reveal sensitive health information.
In other words, alignment itself becomes a privacy breach.
A Paradigm Shift: Align Without Revealing
This paper introduces a fundamentally different approach: a multi-party Private Set Union (PSU) protocol that allows multiple organizations to align their data without revealing whether they share any entities at all .
The key idea is simple but powerful:
- Instead of computing the intersection (what we have in common),
- The system computes a secure union of identifiers,
- While hiding entirely which records overlap across parties.
Figure 3: Illustration of the PSU protocol. All unique IDs across parties form the union dataset used for training; missing features for non-overlapping records are completed with synthetic data.
This transforms privacy from a constraint into an architectural principle.
No raw data is shared. No sensitive relationships are exposed. No trusted third party is required.
Built for the Real World: Noise, Scale, and Security
Beyond privacy, the paper tackles the complexity of real-world data.
Identifiers across organizations are rarely clean. Names have typos, formats differ, fields are split or merged. Traditional exact matching fails under these conditions.
Figure 4: Scheme describing the main steps of the first part of the Diffie-Hellman protocol employed for PSU.
Figure 5: Scheme describing the main steps of the second part of the Diffie-Hellman protocol employed for PSU.
To solve this, the protocol introduces:
- Noise-tolerant matching, capable of aligning imperfect identifiers
- Multi-party scalability, extending beyond the typical two-party limitation
- Lightweight cryptography, reducing communication and computation overhead
- Commutative encryption, ensuring that identical entities can be matched without ever being revealed
Figure 6: Pipeline of the proposed PSU protocol for multi-party VFL.
The result is a system that is not only theoretically sound, but deployable at scale in real environments.
Data Stays Where It Belongs: At the Source
At its core, this work reinforces a crucial principle for the future of AI:
The safest data is the data that never leaves its origin.
All computation happens in a distributed manner. Data remains under the control of each organization. Only encrypted, transformed representations are exchanged.
This architecture delivers:
- Regulatory compliance by design (GDPR, HIPAA, etc.)
- Reduced attack surface, as there is no central data repository
- Full data sovereignty, preserving ownership and control
Security is not an afterthought. It is embedded in every layer of the system.
From Privacy Constraint to Strategic Advantage
What emerges from this work is a broader shift in how we think about data collaboration.
Privacy is no longer a barrier to innovation. It becomes the enabler of collaboration at scale.
Organizations that previously could not share data can now jointly train models. Industries that operated in silos can now build collective intelligence. Sensitive domains such as healthcare, finance, cybersecurity, and defense can finally unlock the value of their data without compromising trust.
Why This Matters Now
As AI systems evolve toward real-world deployment, three factors are becoming decisive:
- Trust
- Security
- Data sovereignty
Accuracy alone is no longer enough.
This paper provides a concrete blueprint for the next generation of AI systems, where:
- Data never leaves its source
- Collaboration happens without exposure
- Privacy is mathematically guaranteed
A New Operating System for Collaborative Intelligence
“Privacy-Preserving Multi-Party Entity Alignment” is not just a technical contribution. It defines a new operational model for AI.
A model where:
- Data remains local
- Intelligence becomes global
- And trust is built into the architecture itself
In this new paradigm, the ability to collaborate without sharing data is not just an advantage. It is the foundation of scalable, secure, and sovereign AI.