Breakthrough in 'Private AI' Tech Announced

The concept of computing on encrypted data, first proposed in 1978, remained largely theoretical for three decades due to practical performance barriers. Craig Gentry's 2009 breakthrough introduced the first plausible Fully Homomorphic Encryption scheme, turning a cryptographic dream into a tangible, albeit slow, reality. The initial implementation took approximately 30 minutes for a single bit operation. Subsequent generations of FHE have focused on reducing this massive computational overhead. The primary bottleneck for AI applications has been the inefficiency of performing matrix multiplication—a foundational operation for neural networks and large language models—on encrypted data. This has kept FHE on the sidelines for most real-world AI workloads. The Gentry-Lee scheme restructures how homomorphic operations handle matrix multiplication, targeting the core of modern AI systems. This new architecture, based on the Ring-Learning with Errors (RLWE) problem, aims to make encrypted computation efficient enough for complex models, moving FHE from a niche security tool to a functional component in the AI stack. This advance underpins the emerging field of "Private AI," enabling AI models to be trained and run on sensitive datasets without ever decrypting them. This is critical for regulated industries like healthcare and finance, which face strict data privacy laws like HIPAA and GDPR, allowing them to leverage AI without exposing confidential information. DESILO, the deep-tech company behind the new scheme, is focused on building infrastructure for this new class of privacy-preserving AI. By making FHE practical, this technology could create a significant moat for companies that can offer AI services with verifiable data privacy, potentially shifting the value in the cloud and AI supply chain toward secure computation providers.