New Encryption Scheme Aims to Make Private AI Practical

The concept of computing on encrypted data, first called a "privacy homomorphism," was proposed in 1978. For over 30 years, however, building a fully functional scheme remained an open problem in cryptography until Craig Gentry's 2009 breakthrough. Gentry's original scheme was the first to prove FHE was possible, using lattice-based cryptography to support both addition and multiplication on encrypted data. This discovery transformed the field from theoretical to practical, earning him the 2022 Gödel Prize for his work. The primary barrier to FHE's widespread use has been performance. Encrypted computations introduce significant overhead, making them up to six orders of magnitude slower than operating on plaintext, a critical bottleneck for real-world applications. This new Gentry-Lee (GL) scheme is dubbed "5th generation," following previous generations that incrementally improved efficiency. Second-generation schemes like BGV and BFV focused on reducing "noise" accumulation during computation, while later schemes like CKKS were optimized for the approximate arithmetic common in machine learning. The GL scheme specifically targets matrix multiplication, the core mathematical workload that dominates modern AI systems and Large Language Models (LLMs). By fundamentally restructuring how this single operation is handled, the new architecture aims to make private AI computation practical. The framework is a joint effort between Gentry, now Chief Scientist at Cornami, and Yongwoo Lee, the Chief Scientist at DESILO. DESILO is a South Korean deep-tech company focused on privacy-enhancing technologies to enable secure data collaboration.