Updates on HEIR, the Homomorphic Encryption Compiler Project

HEIR compiles pre-trained machine-learning models into programs that operate on encrypted data, enabling perfectly private inference. A credit-card fraud detector with a three-layer feed-forward network runs in about 2 seconds on a single-threaded CPU using the Lattigo backend, compared to 0.5 ms for cleartext execution—a 4,000-fold slowdown. The demo is available via a Bazel command that compiles the model and executes the encrypted inference.