Aggregated Zero-Knowledge Proofs Toward Distributed Proof-of-Deep-Learning

Abstract

The recent machine learning requires huge machine resources and is often hard for users with limited resources. Although there are DPT, these are unsuitable for a situation where a trained model should be unrevealed from other users. In this paper, we first propose a new learning method, aggregated zero-knowledge deep learning (AZKDL), whereby even a user with a limited resource contributes to the learning process without revealing its model. Our main idea is to utilize aggregated zero-knowledge proofs where individual zero-knowledge proofs are aggregated into a single proof. Loosely speaking, users generate proofs for their training of parts of models and then aggregate both the models and the proofs to verify the entire models without revealing them. We also prove that AZKDL can detect malicious training. When we conduct experiments to evaluate AZKDL, we identify that even a client with the largest model parameters can finish the computation within a second. Furthermore, we propose the distributed proof-of-deep-learning (DPoDL) that rewards users who contribute to the learning process by applying AZKDL to a mining process of blockchains. DPoDL can detect malicious users by AZKDL.