zk proofs applied to ml (zkml)
tl; dr
- challenges:
- transpile NNs into ZKP circuts (floating-point weigths -> fixed-point arithmetic)
- model size/depth
- ideas:
- model authenticity
- assurance that the ml model is the one that run (e.g. the most accurate one)
- functional commitments allow the prover to establosj that it used a commited model (but no guarantess about the commited model).
- model integrity
- assurance that the same ml algorithm is ran on different data the same way
- attestations
- integrate attestations from external parties
- decentralized inference or traning
- perform ml training in a decentralized way
- proof of personhood
- model authenticity
cool resources
- humanness in the age of ai, by worldcoin
- zk-img: attested images via zk-proofs, d. kang et al.
- checks and balances ml and zk, by a16
- trustless verification of ml, by d. kang
- tachikoma, neural nets for zk proof systems
- zkml, framework for constructing proofs of ml model in zksnarks
- ezkl, deep learning inference in zk-snark
- unraveling zkml, by dr. cathie so