ZK-AI Inference on Ethereum via Boundless & RISC Zero

This project demonstrates verifiable AI inference on-chain using RISC Zero and Boundless. A small neural network trained on MNIST is executed in a zero-knowledge VM (zkVM), and the resulting prediction is verified on the Ethereum blockchain via a smart contract.

The model is run off-chain inside a zkVM, and a zero-knowledge proof (seal) is submitted to an Ethereum smart contract to trustlessly verify the prediction.

🧠 What This Project Does

• Runs a trained neural network (MNIST digit recognizer) in a zkVM
• Produces a ZK proof of the model's prediction
• Submits the prediction and proof to an Ethereum smart contract (MNISTPredictor) for on-chain verification
• Stores the verified prediction on-chain

🚀 Quick Start

1. Install Dependencies

Install RISC Zero and Boundless CLI tooling:

curl -L https://risczero.com/install | bash
rzup install

Install Foundry for smart contract development:

curl -L https://foundry.paradigm.xyz | bash
foundryup

2. Clone the Repository

git clone https://github.com/your-username/zk-ai-mnist-inference
cd zk-ai-mnist-inference

3. Set Environment Variables

Make sure your environment is configured with a valid Ethereum private key and RPC URL (e.g., for Sepolia testnet):

export RPC_URL="https://ethereum-sepolia-rpc.publicnode.com"
export PRIVATE_KEY="your_private_key_here"
export MNIST_PREDICTOR_ADDRESS="deployed_contract_address"

If uploading to IPFS via Pinata:

export PINATA_JWT="your_pinata_jwt_token"

4. Run the Inference App

You can use either a sample MNIST image or provide your own.

Option A: Run with Sample Image

RUST_LOG=info cargo run --bin app -- --sample

Option B: Run with Custom Image File

RUST_LOG=info cargo run --bin app -- --image-file ./input/your_image.txt

Image file format must be a list of 784 integers (28x28 pixels), either as a JSON array, CSV, or whitespace-separated.

🧪 How It Works

1. The guest binary performs inference using hardcoded weights
2. The prediction result is committed to the journal in the zkVM
3. A seal (ZK proof) and journal (output) are submitted via Boundless to Ethereum
4. The MNISTPredictor smart contract verifies the seal and stores the result

🧰 Development

Build Contracts & Guests

forge build
cargo build

Test Contracts and Guest Code

forge test -vvv
cargo test

🛠 Deploying the Smart Contract

Deploy the MNISTPredictor contract to Sepolia:

VERIFIER_ADDRESS="your_verifier_address_here"
forge script contracts/scripts/DeployMNISTPredictor.s.sol --rpc-url ${RPC_URL:?} --broadcast -vv
export MNIST_PREDICTOR_ADDRESS="returned_address_from_deploy"

☁️ Uploading Guest Programs

You can upload the guest binary to IPFS or S3. To use Pinata:

export PINATA_JWT="your_pinata_jwt"

Run the app without the --program-url flag, and it will automatically upload and use the resulting URL:

cargo run --bin app -- --sample

Or, if you already uploaded your guest:

cargo run --bin app -- --sample --program-url https://your.ipfs.link/to/guest

📄 Smart Contracts

• MNISTPredictor.sol: Accepts predictions, verifies ZK proofs, and stores results
• Auto-generated interfaces via alloy are used in the Rust client

✅ Status

• ✅ ZK Inference
• ✅ Seal Generation
• ✅ Ethereum Verification
• ✅ On-Chain Storage
• 🔜 Frontend Interface
• 🔜 NFT badge for correct predictions

🧠 Credits

• RISC Zero
• Boundless Market
• Inspired by boundless-foundry-template

📜 License

Apache-2.0 © 2024 RISC Zero & Contributors