Breakthrough in 'Private AI' Tech
A new, 5th-generation Fully Homomorphic Encryption (FHE) scheme was introduced at the FHE.org 2026 Conference. Developed by DESILO and FHE inventor Craig Gentry, the "GL" scheme promises a major performance breakthrough for processing encrypted data, a key step for private AI.
Fully Homomorphic Encryption (FHE) allows for computations on encrypted data without needing to decrypt it first. The concept, initially termed "privacy homomorphisms," dates back to the 1970s, but it was Craig Gentry's 2009 breakthrough that provided the first plausible construction for a fully homomorphic system. Gentry's initial scheme, while revolutionary, was impractically slow, taking roughly 30 minutes for a single logic gate operation. This performance bottleneck has been the primary challenge for FHE adoption. Subsequent generations of FHE have focused on improving efficiency by reducing "noise" accumulation during computations and optimizing the "bootstrapping" process, a method for resetting this noise. The major hurdle for FHE is its computational intensity, which makes it significantly slower than performing operations on unencrypted data. This performance overhead, along with the complexity of implementing and managing FHE schemes, has limited its widespread use. The development of more efficient schemes is crucial for practical applications. "Private AI" leverages FHE to train and run machine learning models on sensitive data without ever exposing the raw information. This has significant implications for industries like healthcare and finance, where data privacy regulations are stringent. FHE allows AI models to operate on encrypted data, with only the final result being decrypted by the data owner. The security of most modern FHE schemes is based on the mathematical difficulty of lattice problems, which are believed to be resistant to attacks from quantum computers. This positions FHE as a potentially long-term solution for secure computation as quantum technology advances. Previous generations of FHE include schemes like BGV, BFV, GSW, and TFHE, each introducing improvements in noise management, performance, and the types of computations that could be efficiently performed. These advancements have steadily moved FHE from a theoretical possibility to a practical tool for specific privacy-preserving applications.
