ZKML Privacy preserving Credit Loans

Run Info -

Clone our repo.

environment.yml has python install info

Run python ezklCreditLoanPython.py to run the Training of model , generate EZKL proof and verification

Run uvicorn inference_api:app --reload to run the inference.

Open URL - http://127.0.0.1:8000/docs and use json in testing1.json as sample.

🎯 Objective

Our project brings privacy-preserving credit loans app using ZKML (Zero-Knowledge Machine Learning). We wanted to predict a user's credit score — Low, Average, or High — and prove that the classification came from a genuine ML model without revealing any personal data. This score could then be used to decide how much USDC loan to offer the customer.

The neural network model used in your ezklCreditLoanPython.py script is a simple 3-layer fully connected feedforward neural network built using PyTorch.

GITHUB

https://github.com/tenac1337/zkml-encode-credit-score

Testing

ran this inference -

Smart Contract -

• The python notebook generated the Verifier smart contract.
• We created the TokenLender contract which inherits the Verifier smart contract in order to utilise the verification logic
• The submitUnderwritingDecision method on the TokenLender contract takes as input the user address and the proof parameters. Once the proof parameters successfully verify, funds are sent to the user's address.
• This model assumes trust with the underwriter server in providing the correct user address that corresponds to the bundle of private user data they provided to the underwriter. This was a challenge that we did not know how to overcome.

---

🧠 What We Built

• We trained a PyTorch neural network using a custom dataset with features like income, age, education, etc.
• We then used EZKL to convert that model to ONNX and generate zero-knowledge proofs of its predictions.
• Finally, we exposed this via a FastAPI endpoint so anyone can:
  1. Submit input (like age, income),
  2. Get the predicted credit score,
  3. Receive a ZK proof that verifies the prediction was made correctly.

---

⚙️ Tech Stack

• PyTorch for model training
• EZKL for ZKML proof generation
• ONNX for model export
• FastAPI to serve inference and proof as an API
• JSON & joblib for model + preprocessor serialization

---

🛠️ How It Works (Code Flow)

• ezklCreditLoanPython.py:
  • Loads and preprocesses training data.
  • Trains and evaluates the model.
  • Converts to ONNX.
  • Generates + verifies a zero-knowledge proof using EZKL.
• inference_api.py:
  • Takes in live user data.
  • Predicts credit score.
  • Generates a ZK proof of that prediction.
  • Returns both.

---

💡 Real-World Impact

This system can power decentralized loan platforms. Imagine someone applies for a loan and gets a verified credit score, but without ever revealing their salary, age, or education. It’s private, trustworthy, and blockchain-ready.

---

✅ Status

• Working end-to-end!
• Screenshots captured for proof generation & verification.
• My teammate will show how we connect this to USDC lending logic.