An efficient distributive framework for preserving data privacy through block chain

Abstract

Deep Learning Models has gained much attention to perform various artificial intelligence tasks. The accuracy of the models relies on the availability of data. Privacy and auditability has become the major concern for data providers. First issue is the centralised server which may become malicious causing break in privacy. Second is no incentives are given for data providers and trainers. Block chain is the most emerging innovation as of late. Decentralised connectivity of block chains gives another approach to interface information without the overheads of security, trust and controls. To address the above issues we propose an algorithm where clients send the model to the block chain for training where the honest trainers are incentivized for training, sharing weights. The weights are averaged; parameters are updated by a smart contract that resides on block chain which guarantees privacy and audit ability.