Circom-MPC Summary

Circom-MPC - a project, that allows to port Circom zk-circuits to MPC circuits and to run it with a picked backend.

This topic covers two tools:

circom-2-arithc - Fork of Circom; compiles circuits to Bristol Fashion format
circom-mp-spdz - transpiler from Bristol format to .mpc files to run with MP-SPDZ backend; might be changed for other backends (s.a. mpz) too

circom-2-arithc

Most up to date info can be found in README.md

Aim of the project - to have a DSL that developers are familiar with - for MPC, + to port currently available circuits/protocols with no need to change the code.

Current state - supports Circom 2.1.0; has some minor bugs & not optimized (for big circuits, esp. with many loops, takes much time to compile).

Circom-2-arithc can compile circuits to both Arithmetic circuits & Boolean circuits (specified by the flag); also it has minimal support of types (e.g. MP-SPDZ's sint | sfloat) - the info is added to output file and e.g. can be used on a .mpc file generation stage.

circom-mp-spdz

Set of scripts to transpile (from Bristol to .mpc), run (with MP-SPDZ), benchmark (including remote machines benchmarking) and check accuracy of the output (used for ML circuits).

ML circuits are used as a main use-case and example for the project.
The circuits were taken from [socathie/circomlib-ml])(https://github.com/socathie/circomlib-ml), they can be ported to MPC with no changes.
As a result we executed, benchmarked and checked the accuracy for all circuits. More info on results can be found in BENCHMARK.md and ACCURACY.md.

An in-depth review of circomlib-ml-mpc can be found here. This file also contains the instruction on how we can benchmark/test the circuits.

Short overview for Keras2Circom-MPC

As a part of Circom-MPC project we also were able to run synthetic and real-world ML model in MPC.
For that we used Keras2Circom. Keras2Circom allows to compile Keras model to Circom circuit - thus, having the compiler we can run Keras2Circom generated circuit in MPC.

We generated two circuits with Keras2Circom:

Synthetic model - a model that does nothing useful: it only combines many layers; it's created for testing purposes
Real-World model - real-world (useful) ML model, taken from keras2circom example. We were able to compile it, and it's possible to run it, but it takes ~30 minutes to compile on M3 Max (due to non optimized compiler) and many hours to run (more than 107000 gates and wires)

Open questions

Should we continue working on Circom-MPC? Is it really useful for devs? Is there a demand for MPC apps creation?
What features should we add to the compiler and how can we optimize it?
What other applications should we explore and build with the given tooling?