Google open-sources HEIR homomorphic encryption compiler for AI inference on encrypted data
Google open-sourced HEIR (Homomorphic Encryption Intermediate Representation) compiler on August 14, enabling pre-trained models to perform inference directly on fully homomorphic encrypted data for healthcare, finance, and legal data residency.
On August 14, Google released and open-sourced the HEIR (Homomorphic Encryption Intermediate Representation) compiler as part of its Private Computing Toolkit. Built on MLIR, the compiler converts pre-trained ML models to versions that run on fully homomorphic encrypted (FHE) data, allowing cloud services to perform inference without ever seeing plaintext.
HEIR aims to let non-cryptographers add encrypted inference to production without assembling a dedicated crypto team. Google partnered with hardware accelerator vendors Belfort, Niobium, Cornami, and Optalysys, and collaborated with Georgia Tech, Carnegie Mellon, UC Santa Barbara, Purdue, Tsinghua, and the University of Edinburgh to position HEIR as a research platform.
Demos released include encrypted credit card fraud detection, network threat detection, and a hotword detector. The core significance of HEIR is closing the gap that historically made FHE impractical for production ML. Google security blog author Jeremy Kun noted that FHE inference compute cost is falling rapidly, shifting the privacy/capability trade-off from "impossible" to "how much will you pay."
For healthcare, finance, and legal data residency scenarios, FHE inference removes the need to send plaintext to cloud hosts. Independent benchmarks of FHE overhead versus plaintext inference speed will be the key metric to watch.