Google released HEIR on August 14, an open-source MLIR-based compiler that converts pretrained models to run inference on encrypted inputs, so a server can compute a prediction without ever seeing the data. You annotate which Python types are secret and the compiler handles packing, parameter selection and arithmetization.
Read the full story: Google's HEIR Compiler Runs AI On Encrypted Data →
Transcript
Google just open sourced a compiler that lets a server run an AI model on data it cannot read. It is called HEIR. The input arrives encrypted, the computation happens without ever decrypting it, and only the client can open the result. Fully homomorphic encryption has been possible since two thousand nine. The math was never the problem. The problem was that making a real model run under it meant hand tuning ciphertext layouts and picking encryption parameters, which took specialists months. HEIR turns that into a compiler pass. You write Python, you annotate which values are secret, and it handles the rest. Google shipped four working demos, including card fraud detection and voice assistant hotword detection. Here is the catch. The announcement publishes no latency figures at all. For a technology whose entire history is about overhead, that is the number everyone needs. Go measure it yourself before you promise it to anyone.