Secular: A Decentralized Blockchain-based Data Privacy-preserving Model Training Platform

The main actors for productizing machine learning models are machine learning model builders. Commonly, they aim to train their models on a large enough, suitable, convenient, and realistic data set to get a model of expected and satisfactory accuracy. However, one of their biggest challenges is how to maintain the security and privacy of the given dataset. Besides, Blockchain is one of the most discussed technologies of our time. It can be described as a radical new approach to delivering trust and confidence over exchanges of value without relying on a trusted third party. Hence, this article aims to propose Secular, a decentralized Blockchain-based data privacy-preserving model training platform. The main objective of the proposed platform is to guarantee data security and privacy while allowing model builders to upload their models for training as well as sharing the required data. We present, in this article, a high-level architecture for the data providing, model uploading, model validation, and model training phases. Finally, the Software Requirements Specification (SRS) and a prototype of Secular are given.

I. Introduction

Today, more than ever, a collective approach and centralized platform has become of vital importance to facilitate our collaborative research efforts in various scientific domains [1]. Unfortunately, most data owners are disinclined to publicly disclose their data. Without a doubt, it is better to use as much data as possible when Machine Learning (ML) techniques are utilized to extract patterns or models, i.e., knowledge, [2]. For instance, if multiple parties are gathering or collecting data for a specific domain, it is preferable to share all their data for knowledge discovery. However, such sharing will not be amenable unless the distinct parties could establish trust relationships among themselves [3]. We are interested in a method to provide a secure private platform where ML model builders can upload their models for training, as well as sharing the required data in a secured and privacy-preserving manner.