Expanding our Fully Homomorphic Encryption offering- Google Developers Blog

What this post added

Introduced new tools to expand Fully Homomorphic Encryption (FHE) capabilities for video files and other complex data. This includes a TensorFlow-to-FHE compiler for trained ML models, the Jaxite software crypto library for running FHE on hardware accelerators (GPUs, TPUs), and the Homomorphic Encryption Intermediate Representation (HEIR) open-source compiler toolchain for FHE program interoperability. The TensorFlow-to-FHE compiler can compile TensorFlow Lite models to FHE, achieving private inference in 16 seconds for a 3-layer neural network. Jaxite leverages JAX for FHE computations on accelerators. HEIR is built on MLIR to lower barriers to privacy engineering and research.