FAIRZK: A Scalable System to Prove Machine Learning Fairness in Zero-Knowledge

Overview

This repository contains the code for FairZK, accepted at IEEE S&P 2025.

FairZK designs a specialized SNARK protocol to prove the fairness bound of a DNN model from its parameters and some aggregated information of the input without doing real inference.

Setup

1. Install Rust: Follow the instructions on Rust Installation.

2. Verify Installation: Post-installation, ensure everything is set up correctly with:

   cargo --version
   rustup --version

3. Use the Nightly Toolchain:

   rustup default nightly

Run the Code

To run the example /fairzk/examples/total.rs, which is a 3-layer DNN on Adult dataset:

cargo run --example total --release

expected output example:

row_num: 128, col_num: 37
row_num: 128, col_num: 128
row_num: 1, col_num: 128
row_num: 45222, col_num: 37

which presents shapes of weight matrices

construct lookup instance: 13 ms
commit oracles: 81 ms
commit lookup first oracles: 81 ms
commit lookup second oracles: 79 ms
prove infairence sumcheck: 0 ms
prove spectral norm sumcheck: 2 ms
prove l2 norm sumcheck: 0 ms
prove bound sumcheck: 0 ms
prove total snark lookup sumcheck: 29 ms
prove oracles: 253 ms

== Prove Timing (in ms) ==
Construct:          13 ms
Commit Oracles:     241 ms
Prove Oracles:      253 ms
Prove Sumcheck:     31 ms

which presents the components of the prover time

verify infairence sumcheck: 0 ms
verify spectral norm sumcheck: 0 ms
verify l2 norm sumcheck: 0 ms
verify bound sumcheck: 0 ms
verify lookup sumcheck: 0 ms
verify oracle: 184 ms

== Verify Timing (in ms) ==
Verify Oracles:     184 ms
Verify Sumcheck:    0 ms

which presents the components of the verifier time

proof size total 349 MB

which presents the proof size

total success

which presents whether the verifier accepts the proof

fairzk/piop contains PIOPs for different components, such as spectral norm. fairzk/piop/total combines all the components.

fairzk/snark contains usable SNARKs by compiling PIOPs with PCS.

models contains quantized parameters of multiple models. You can modify the file paths in /fairzk/examples/total.rs to evaluate the peformance on different models.

Acknowledgements

The implementation of algebra is based on the following libraries: arkwork, plonky3, primus-fhe. The implementation of fairzk/sumcheck is based on arkwork.