Privacy-Preserving Collaborative Deep Learning Using Verifiable Multi-Secret Sharing Scheme

Abstract

Collaborative deep learning is an approach used to handle the amount of training data needed to build a better deep learning model. In collaborative deep learning, the central server collects user data and run the deep learning algorithm to get more accurate models. However, centralized training data collection can cause serious like privacy leakage problems and damage to the integrity of training data. In this paper, we introduce the privacy-preserving collaborative deep learning model using verifiable (k, t, n) multi-secret sharing based on the Elliptic Curve Diffie Hellman and SHA3-256 as a hash function. Where all training data will be formed into n shares using a session key generated from the private key and public key Elliptic Curve Diffie Hellman to protect the privacy and avoid all training data using SHA3-256 for the verification process before sending to the server. The test results show the integrity of damaged training data and colluding participants can be verified. In addition, the accuracy of the model produced using or without using Verifiable Multi-Secret Sharing Scheme has the same value. Therefore proposed model can protect the privacy and integrity of training data and maintain the accuracy of the deep learning model.