Google announced HEIR (Homomorphic Encryption Intermediate Representation) on August 14, an open-source compiler toolchain that lets AI models run inference directly on encrypted data, without ever decrypting it on the server.
The tool targets a long-standing trade-off: strong encryption protects user data but usually blocks providers from offering data-dependent features like recommendations or fraud checks. HEIR converts pre-trained models so they operate on encrypted inputs instead, with Google describing the goal as a "one-click" path for non-experts to add encrypted inference to production apps. Google demonstrated the approach on four applications: a recommendation model, credit-card fraud detection, network intrusion detection, and hotword detection for voice assistants.
The project already has hardware-accelerator partners including Belfort Labs, Niobium, Cornami, and Optalysys, plus academic collaborators such as Georgia Tech, Carnegie Mellon, and Tsinghua University. Four peer-reviewed papers have been built on HEIR so far, with more in preparation.