Google publishes HEIR, a compiler that runs model inference on homomorphically encrypted inputs

Sources: primary, discussion
Summary: HEIR is open source and converts pre-trained models rather than requiring a program to be rewritten for homomorphic encryption by hand. Google names four compiled applications: a deep learning recommendation model, credit card fraud detection, the Kitsune network anomaly detector, and a hotword detector, and states the reported latency figures are single-threaded CPU. It names hardware accelerator partners and states those latency benefits are to be demonstrated later rather than now.
Why it matters: Converting a program to homomorphic encryption by hand needs a team of cryptographers, and a compiler moves that cost to models teams already have.