How Google is Making Private AI Practical with Homomorphic Encryption
Google is open-sourcing HEIR, an MLIR-based compiler toolchain that converts pretrained models to run inference on encrypted inputs. The post demonstrates recommendation, fraud detection, encrypted-traffic intrusion detection, and hotword models, while acknowledging that homomorphic encryption still carries substantial overhead.
The HN discussion focused on that caveat: commenters with privacy-preserving-ML experience questioned commercial viability at roughly 10³ overhead, while others clarified that FHE ciphertexts remain computationally indistinguishable from noise and debated whether the provider can prove it ran the intended model.