Paper 2017/1038

Embedded Proofs for Verifiable Neural Networks

Hervé Chabanne, Julien Keuffer, and Refik Molva

Abstract

The increasing use of machine learning algorithms to deal with large amount of data and the expertise required by these algorithms lead users to outsource machine learning services. This raises a trust issue about their result when executed in an untrusted environment. Verifiable computing (VC) tackles this issue and provides computational integrity for an outsourced computation, although the bottleneck of state of the art VC protocols is the prover time. In this paper, we design a VC protocol tailored to verify a sequence of operations for which existing VC schemes do not perform well on \emph{all} the operations. We thus suggest a technique to compose several specialized and efficient VC schemes with Parno et al.'s general purpose VC protocol Pinocchio, by integrating the verification of the proofs generated by these specialized schemes as a function that is part of the sequence of operations verified using Pinocchio. The resulting scheme keeps Pinocchio's property while being more efficient for the prover. Our scheme relies on the underlying cryptographic assumptions of the composed protocols for correctness and soundness.