Getting Started with HEAL

HEAL is an interface that hardware teams implement to connect with our fully homomorphic encryption (FHE) software stack, giving access to run real AI models on encrypted data.

We provide unit tests and runtime tools to help verify correctness and measure performance during development.

Workflow Overview

Follow these steps to integrate with HEAL and run AI workloads using your hardware implementation:

1

Understand the Architecture

Learn how HEAL connects hardware to the FHE software stack and familiarize yourself with core concepts.

• HEAL uses JSON transcripts containing sequences of commands and encrypted dummy inputs. These transcripts specify the order of function execution and required inputs.

• A runtime script processes these transcripts sequentially, executing each specified action.

• We work with tensors and use the CKKS encryption algorithm. Currently, we use CUDA for parallel processing, but HEAL is built flexibly, allowing any hardware that meets our requirements to easily plug into our system.

2

Implement Functions

Check our specification guide for instructions on implementing HEAL functions.

• You don't need to implement every function - you can choose which ones work best for your hardware.

• We can also work with you to add new functions that match your hardware's capabilities.

3

Run Unit Tests

Validate each function with pre-built test packs from our GitHub repository.

4

Execute Runtime

Run your full implementation on encrypted AI inputs to measure actual performance and identify areas for optimization.

➀ Understand how HEAL works and the terms we use.

➁ Learn how to implement functions that HEAL expects.

➂ Run unit tests using our GitHub validation packs.

➃ Run your hardware-implemented functions on a real AI model through our FHE runtime.