zkML Proof Generation

Introduction

zkML proof generation involves creating a single zero-knowledge proof (ZKP) of a given output from the zk-circuit of a committed ML model. The resultant proof allows any third-party verifier (with access to the required ezkl collaterals) to trustlessly verify whether the claimed output did indeed come from the claimed zk-circuit of a given model.

After the one-time setup of ezkl has been successfully completed, proof generation is relatively straight-forward and takes much less time compared to the setup.

Overview

At a high level, proof generation involves the following steps:

• Generate a witness file through ezkl.gen_witness()

• Perform a mock run through ezkl.mock() as a sanity check to confirm the correctness of the witness file

• Generate the proof through ezkl.prove()

Further technical details and a step-by-step guide can be found here.

Benchmarks

Some ezkl benchmarks related to proof generation across a range of model sizes on a local laptop are:

# of parameters	Proof Generation	Proof Verification	Proof file size
10,497	1.484s	0.031s	24.9kB
17,386	2.897s	0.059s	24.6kB
45,817	5.012s	0.093s	24.4kB
118,721	12.83s	0.187s	38kB
197,681	24.689s	0.337s	38kB
479,873	49.106s	0.691s	38kB
1,041,225	147.472s (02:27)	2.342s	38kB
2,010,929	212.956s (03:33)	2.89s	38kB

Proof Submissions

From Spectral App's v0.2.0 onwards:

• all model predictions submitted during the model validation phase will need to be accompanied by their respective ZKPs

• During the Consumption Window, Validators may regularly request ZKPs for random predictions to ensure continuous transparency and integrity of the platform