HYDRA: Pipelineable Interactive Arguments of Knowledge for Verifiable Neural Networks

We present Hydra, a novel verifiable computation system for neural networks. In this environment, a client (verifier) requests a server (prover) to evaluate a computation represented as an arithmetic circuit. The server returns the result of said computation, as well as a proof that attains to its correctness. Hydra introduces an interactive argument scheme protocol geared towards the efficient pipelining of general arithmetic circuit verification, where layers of the circuit can be proven asynchronously. Compared to non-interactive SNARKs which rely on either knowledge type assumptions or the Random Oracle model and theoretical non-interactive arguments based on standard assumptions that are not useful in practice, Hydra achieves a sweet spot with a practical approach. From standard assumptions, we collapse the round complexity to polylogarithmic to the width of the circuit. We demonstrate the effectiveness of this protocol in general cloud computing environments, and also apply this protocol in the specific context of cloud-based deep neural network verification. We propose a new interactive neural network quantization algorithm to convert a neural network into a provable arithmetic circuit representation, and leverage Hydra's pipelined approach to efficiently verify neural networks composed of many layers in both inference and training. We obtain protocol time efficiency improvements of up to 34.8 X for general computation while also reaching state of the art accuracy (99.5%) on MNIST for our neural network framework.