Towards Scalable ZKML "training" via Prior Fitted Networks

What if we could prove what training data was used to make predictions, in the same way we prove inference?

Overview

• giza_tabpfn_demo.ipynb: Contains end-end example using Giza Actions and Giza Model to perform perform end-end supervised classification on a tabular datset from OpenML.
• Within the zk_tabpfn directory, the TabPFN REPO has been ported in :
• transformer_prediction_interface.py: Has been slightly modified to work with Giza Actions.

Requirements

• Python 3.11
• Conda, pip (or Poetry)

Get Started

$ conda activate env
$ pip install -e .

or with poetry

$ poetry shell
$ poetry install

Learn More

Note: This project worked from the Giza's MNIST tutorial. Please be aware that certain steps, such as transpiling the model and deploying the generated model on Giza Plateform, are required between action executions. For a more comprehensive understanding, refer to the tutorial.

Explore more about the Giza Actions SDK here.