Privacy rules increasingly push AI workloads toward encrypted data, yet the tools to run models on ciphertext remain a niche research skill. Google’s answer is a compiler called HEIR that accepts ordinary pre-trained models and re-targets them for encrypted execution.
The project’s full name, Homomorphic Encryption Intermediate Representation, describes what it is: a middle layer between a model and the math that keeps data hidden. Teams that already have a working network for plaintext inputs can compile the same weights to run over encrypted inputs, so a server can process sensitive prompts without ever seeing them in the clear.
The practical pitch is one-click deployment. Encrypted inference has historically meant hand-built pipelines and painful performance trade-offs; by accepting standard models, HEIR removes the re-engineering step that kept the technique inside research labs.
For banks, hospitals, and government agencies that hold data too sensitive to hand to a cloud provider, that changes the cost-benefit math of using outside AI at all. The compiler is open source, so the same capability is available to any team willing to pay the latency price.
