Desilo, a privacy-enhancing technologies (PET) company, said on April 28 it has launched commercial software that implements the "GL Scheme (Gentry-Lee Scheme)," which it describes as a fifth-generation fully homomorphic encryption (FHE) technology.

The company said the practical use of "Private AI," which trains and runs inference on AI models while data remains encrypted without the risk of leaks, has moved closer.

Homomorphic encryption is a cryptographic technology that can perform computations directly on encrypted data without decrypting the original data, the company said. It said existing homomorphic encryption has faced speed limits in handling matrix multiplication, a key operation in modern AI computing, because of massive computational costs.

The GL Scheme is a fifth-generation homomorphic encryption algorithm jointly developed by Craig Gentry (크레이그 젠트리), the inventor of homomorphic encryption and a chief scientist at Cornami, and Yongwoo Lee (이용우), a chief scientist at Desilo. It is designed to improve the efficiency of matrix multiplication, it said.

The "DESILO FHE Library" introduced by Desilo is commercial software that implements fifth-generation homomorphic encryption for immediate use in industrial settings.

It is optimised for high speed in both general computing environments using CPUs and in high-performance graphics processing unit (GPU) environments, the company said. It also provides a base Python environment, which it said is the most widely used programming language in the AI industry.

Yongwoo Lee said, "In modern AI systems, matrix multiplication is the core workload that accounts for the largest share." He added, "Desilo software using the GL Scheme improves how this key operation is handled in a homomorphic encryption environment, and will provide a technological foundation to bring private AI, which had remained theoretical, into real industrial environments."