JSTprove is a command-line toolkit for generating zero-knowledge proofs (ZKPs) of AI inference on ONNX models. It enables private verification of ML computations via a streamlined pipeline from ONNX input to ZKP generation, leveraging Polyhedra Network's Expander and ECC.

How It Works

A Python frontend orchestrates the pipeline: ONNX models are quantized, compiled into arithmetic circuits using Expander Compiler Collection (ECC), and then processed by Polyhedra Network's Expander (GKR/sum-check prover) for ZKP generation. The design emphasizes explicit control, reproducibility, and a user-friendly CLI, supporting core ML ops like Conv2D, GEMM, ReLU, and MaxPool2D, with optimizations like circuit fusion.

Quick Start & Requirements

• Installation: Recommended via uv tool install JSTprove. Development requires cloning the repo, Rust (nightly), system dependencies (OpenMPI, clang/llvm), cloning Expander as ./Expander, and uv sync.
• Python: 3.10–3.12.
• Dependencies: UV, OpenMPI, Rust nightly, clang/llvm.
• Demo: LeNet demo illustrates compile, witness, prove, verify CLI workflow.
• Docs: docs/cli.md, docs/troubleshooting.md, docs/CONTRIBUTING.md.

Highlighted Details

• Generates ZKPs for AI inference on ONNX models.
• Utilizes Polyhedra Network's Expander and ECC.
• Supports Conv2D, GEMM/MatMul, ReLU, MaxPool2D.
• Emphasizes explicit paths, reproducibility, and deterministic debugging.
• Includes quantization and circuit fusion.

Maintenance & Community

No specific maintainer or community details (Discord/Slack) are provided. Contributions are welcomed but advised with caution.

Licensing & Compatibility

No explicit license is stated. A prominent disclaimer warns that JSTprove is experimental, unaudited, provided "as-is," and strongly discouraged for production use due to potential bugs and security risks. Users assume all responsibility.

Limitations & Caveats

JSTprove is experimental and unaudited, unsuitable for production. It may contain bugs/vulnerabilities and is subject to breaking changes. Python 3.13 is incompatible. Setup requires careful installation of Rust nightly, OpenMPI, and build tools.