Privacy-Preserving Machine Learning techniques for Horizontal Distributed Data: A Survey

Abstract

Privacy-preserving machine learning is the development and deployment of Machine learning (ML) models whereby the data should be shielded from personal privacy concerns during model training. This is the main problem in a distributed system when data is spread over several sites, possibly exposing sensitive information while training the data, particularly in the healthcare and finance industries. There are several ways to address these privacy issues, including differential privacy, secure multi-party computation (SMPC), homomorphic encryption (HE), and federated learning. Techniques for deep learning and machine learning (ML) have a lot of promise for raising productivity. To get decent results, however, the data used to train machine learning models must be of very high quality. Only when there is a large amount of flawless data provided for training can any machine learning algorithm function exceptionally well. In this paper, we offer strategies and provide detailed survey and analysis of privacy-preserving ML techniques such as HE, Multi-party Computation, Federated Learning and Differential Privacy. The proposed work includes analysis of existing techniques and information on the design and implementation of various PPML protocols. We also cover the benefit of privacy during computation in real time applications, which, because to its distributed, secure, and private nature, has the ability to address the security issues raised above.