Commit to model inputs rather than provide them to the Verifier #201
Description
Goal is to be able to hide model's inputs from the verifier.
Current behavior
The model's inputs are given in clear text to the Verifier, together with the model outputs.
During backpropagation, each time the verifier encounters an Input node, it asserts that claimed evaluation at random point matches the evaluation of prover-provided input Tensor. If so, the claimed output was indeed recovered by executing the model on prover-provided inputs.
Proposed
The prover commits to model inputs, and sends the commitments as part of the proof.
During backpropagation, each time the verifier encounters an Input node, it saves its claimed opening as an Input(usize) CommittedPolynomial variant, which will be verified later during CommittedPolynomial's opening verification, together with any other committed polynomial.