Development of Privacy Preserving Machine Learning Techniques Using Secure Multi-Party Computation

Title

Development of Privacy Preserving Machine Learning Techniques Using Secure Multi-Party Computation

Creator

Tiwari, Kapil

Contributor

George, Jossy P

Description

Machine learning (ML) has brought about a paradigm shift in insight generation across various domains, including healthcare, finance, and pharma, by leveraging historical data. However, the effectiveness of ML solutions hinges on the seamless collaboration between data owners, model owners, and ML clients while ensuring that privacy concerns are meticulously addressed. Unfortunately, existing privacy-preserving solutions have not been able to offer efficient and confidential ML training and inference. This has led to an increased focus on Privacy-Preserving Machine Learning (PPML), which has become a flourishing area of research aimed at safeguarding the privacy of machine learning stakeholders. In this regard, the present research introduces novel techniques for private ML inference and training of models using Secure Multi-Party Computation (SMPC) and Differential Privacy (DP) methods on horizontally and vertically partitioned datasets. The proposed techniques are implemented using Python with open-source libraries such as SyMPC and PyDP to ensure confidential inference and model protection. The findings across various parameters illustrate the effectiveness of the suggested techniques in addressing the privacy concerns of model owners and inference clients, with no significant impact on accuracy and a linear influence on performance as the privacy parameters, such as secure nodes count within the SMPC cluster. are increased. Furthermore, the privacy gain is substantiated by information privacy measures such as Mutual Information and KL-Divergence across different privacy budgets, which demonstrate empirically that privacy can be preserved with high ML accuracy and minimal performance cost.

Source

Author's Submission

Date

2023-01-01

Publisher

Christ(Deemed to be University)

Subject

Computer Science

Rights

Open Access

Relation

61000306

Format

PDF

Language

English

Type

PhD

Identifier

http://hdl.handle.net/10603/548041