Multi-key fully homomorphic encryption, or multi-key FHE, allows several organisations to run a joint calculation on encrypted data while each organisation keeps its own encryption key.
A CRYPTO 2026 paper removes a size constraint that affected every previous multi-key FHE design. The encrypted packages and keys no longer need to grow as more participants join.
The result is a cryptographic construction. It has not been tested as production banking software.
Key facts
- Multi-key FHE supports calculations across data encrypted by different organisations.
- Participants do not need to pool their original data or use one shared encryption key.
- The CRYPTO 2026 result keeps encrypted package and key sizes independent of the number of users.
- The research applies to joint computation, including some privacy-preserving AI use cases.
- Processing cost, implementation security and institutional governance still require further work.
What is multi-key fully homomorphic encryption?
Fully homomorphic encryption, commonly shortened to FHE, allows a computer to calculate using encrypted information.
Most computing systems must decrypt information before using it. The data becomes readable during processing, even if it was encrypted while stored or transferred.
FHE keeps the information encrypted during the calculation. The system returns an encrypted result that only an authorised party can unlock.
Multi-key FHE allows the calculation to include data protected by different encryption keys. Each organisation keeps control of its own key instead of placing all participating data under one shared key.
Why would banks use multi-key FHE?
Banks, payment providers and asset managers observe different financial signals.
One bank may detect an unusual payment pattern. Another may identify suspicious account activity. A payment provider may see related transactions across several merchants.
Joint analysis could identify fraud or financial crime that would be difficult to detect from one institution’s records alone. It could also support market-risk analysis across several firms.
The institutions cannot simply copy customer records, transaction histories or trading positions into a common database. Legal restrictions, commercial confidentiality and cybersecurity risks limit how this information can be shared.
A central database also gives its operator access to a large collection of sensitive information.
Multi-key FHE removes the need for one shared key. Each institution encrypts its own data before contributing it to the calculation. The system operator processes encrypted inputs and does not need the original records.
What stopped multi-key FHE from scaling?
Every previous multi-key FHE construction had a size problem.
Parts of the encryption system became larger as the number of users increased. More participants meant larger encrypted packages and more key material.
This made larger institutional networks harder to operate. A system involving five organisations required fewer cryptographic resources than one involving fifty, even before accounting for the calculation itself.
Researchers had not found a way around this constraint under standard security assumptions.
What did the CRYPTO 2026 paper prove?
Researchers Abtin Afshar and Rishab Goyal presented the first multi-key FHE construction in which the encrypted package, public key and secret key sizes remain independent of the number of users.
Their paper was approved by the International Association for Cryptologic Research on 15 August 2026 and published in connection with CRYPTO 2026.
Under this construction, adding more organisations does not automatically increase the size of the encrypted packages or keys.
The processing cost may still increase. Running a calculation across more data can require more computing resources. The paper addresses the size of the encrypted objects and keys, not every cost associated with the system.
The construction also supports joint decryption in a single round. Each participating organisation contributes a decryption share without revealing its complete secret key.
This allows a group to control access to the final result. The rules can require participation from several institutions before the result is unlocked.
Where could multi-key FHE be used in financial AI?
Multi-key FHE may support AI systems that analyse information from several institutions without collecting the original records in one place.
Possible uses include:
- Comparing encrypted fraud indicators across banks
- Detecting suspicious payment activity across separate networks
- Calculating shared market-risk measures without exposing individual positions
- Examining financial-crime signals without pooling complete customer files
These are proposed applications, not deployments reported in the CRYPTO 2026 paper.
The research addresses one component of the security design. It allows each institution to retain its own encryption key while participating in a joint calculation.
An operational system would still need to define who can submit data, who can run the model, who can access the result and what happens when an institution leaves the group.
Can an AI system process data without seeing it?
FHE allows software to process encrypted information without first converting it into readable data.
An AI model used within an FHE system could receive encrypted inputs and produce an encrypted output. The company operating the model would not need direct access to the original information.
Other information can still be exposed if the system is poorly designed. Access records, usage patterns, metadata and final outputs may reveal sensitive details.
Institutions must protect these elements separately. They must also set rules for requesting calculations and approving access to results.
FHE protects data during computation. Identity management, audit records, contracts and regulatory controls remain necessary.
Is multi-key FHE ready for banks?
The CRYPTO 2026 paper presents a mathematical construction. It does not report a banking deployment, commercial implementation or application-level performance test.
Its security relies on the Learning with Errors assumption, which is widely studied in modern cryptography.
Processing performance remains a practical constraint for FHE systems.
A separate 2026 research system called Cachemir tested encrypted inference using the Llama-3-8B language model. The researchers reported CPU speed improvements of 48.83 times and 67.16 times over two earlier approaches.
The GPU implementation still required less than 100 seconds to generate one output token.
Cachemir is not a benchmark of the Afshar and Goyal construction. It provides context on the current cost of running larger AI models with encrypted data.
Implementation work is now required to measure the new multi-key construction against real workloads. The code would also require independent security review before institutional use.
What does multi-key FHE leave unresolved?
Multi-key FHE protects information during a calculation. It does not assess whether the data or model should be used.
It does not:
- Correct inaccurate or incomplete data
- Test an AI model for bias or poor decisions
- Establish legal permission to use the information
- Prevent an authorised output from revealing sensitive details
- Protect every part of the surrounding software and hardware
- Assign responsibility between participating institutions
A bank considering this technology would still need performance testing, cybersecurity controls, audit records, governance procedures and regulatory approval.
What should financial institutions take from the research?
The paper removes a specific constraint in multi-key FHE. A larger group no longer requires larger encrypted packages and keys solely because it has more participants.
This improves the design options for institutions that want to calculate across private data while maintaining separate control.
A multi-institution system must distinguish four responsibilities:
Ownership of the data Control of the encryption keys Operation of the calculation Authority to unlock the result
Multi-key FHE can enforce parts of this separation. Legal agreements and operating controls must cover the rest.
e23 applies a rails-before-products approach to financial infrastructure. For privacy-preserving AI, those rails include defined ownership, separate keys, controlled computation and recorded approval.
The CRYPTO 2026 result addresses the size of the encrypted packages and keys. Performance, implementation and governance remain open work.
Frequently Asked Questions (FAQ)
What does FHE stand for?
FHE stands for fully homomorphic encryption. It allows a computer to calculate using encrypted information without first decrypting it.
How is multi-key FHE different from standard FHE?
Standard FHE commonly uses data protected by one encryption key. Multi-key FHE can calculate across data protected by different keys belonging to different organisations.
Does multi-key FHE require banks to share raw data?
No. Each bank can encrypt its information before contributing it to the calculation. The original records do not need to be placed in a shared database.
Can the AI provider read the bank data?
A correctly implemented FHE system does not require the model operator to read the original data. The inputs remain encrypted during processing.
Is the CRYPTO 2026 construction available as banking software?
No. The paper presents a cryptographic construction. Implementation, testing and independent review are still required.
Is multi-key FHE fast enough for large AI models?
Current FHE systems are much slower than conventional computing. The CRYPTO 2026 paper solves a size constraint. It does not establish production-level performance for large AI models.
Does multi-key FHE replace data-governance requirements?
No. Institutions still need contracts, identity controls, audit systems, regulatory approval and clear responsibility for the data and results.
Source
- Fully-Succinct Multi-Key FHE & Rate-1 Simulatable Threshold Decryption from LWE, Abtin Afshar and Rishab Goyal, IACR ePrint 2026/1683, CRYPTO 2026. Approved 15 August 2026.
- Cachemir: Fully Homomorphic Encrypted Inference of Generative Large Language Model with KV Cache, Ye Yu et al., arXiv:2602.11470. Submitted 12 February 2026.