Google's project team has released details regarding HEIR, a Fully Homomorphic Encryption (FHE) compiler. HEIR is a compiler that transforms input programs into programs capable of performing direct operations on encrypted data. This technology allows computers to continue processing without ever knowing the contents of the input data, intermediate values, or output data, all without decryption.
HEIR, a compiler enabling computation on encrypted data, shows 4,000x latency in ML model inference
This article is a translation. Read the Japanese original
HEIR also supports the compilation of pre-trained machine learning models, enabling the provision of inference services while maintaining complete privacy. The project roadmap and detailed mechanisms are available in the official documentation and other sources.
As a proof of concept, an example using a credit card fraud detection model was presented. This model consists of a three-layer feedforward network. In a single-threaded CPU environment, inference using encrypted data took approximately 2 seconds. In contrast, the execution time using plaintext data was 0.5 milliseconds.
The processing time due to encryption represents a latency of approximately 4,000x compared to plaintext. This experiment included two matrix-vector multiplications and two sigmoid function evaluations.
Source: Updates on HEIR, the Homomorphic Encryption Compiler Project (Hacker News Frontpage, 2026-09-05)