[D] The engineering overhead of Verifiable ML: Why GKR + Hyrax for on-device ZK-ML?

About this article

The idea of ​​"Privacy-Preserving AI" usually stops at local inference. You run a model on a phone, and the data stays there. But things get complicated when you need to prove to a third party that the output was actually generated by a specific, untampered model without revealing the input data. I’ve been looking into the recently open-sourced Remainder prover (the system Tools for Humanity uses for World). From an ML engineering perspective, the choice of a GKR (Goldwasser-Kalai-Rothblum) +...