DPLE: a Privacy-Enhanced and Straggler-Resilient Distributed Learning Framework for Smart Cloud

EasyChair Preprint 12740, version 1

Abstract

In the smart cloud environment, distributed learning faces privacy and straggler issues. Lagrange coded computing can alleviate these concerns to some extent. However, when the number of honest but curious nodes exceeds a certain threshold, or there exists outside eavesdroppers, the privacy of the system will be threatened. To address this challenge, we propose a differentially private Lagrange encoding distributed learning framework, named DPLE. Firstly, we utilize Lagrange encoding to hide the raw data and inject redundancy, thereby enhancing privacy protection and resilience against stragglers. Additionally, artificial noise will be injected into local computation results, further securing sensitive information against leakage. Moreover, we conduct theoretical analyses to determine the variance of artificial noise required to achieve a certain level of privacy protection within this framework. Through experiments, we validate the effectiveness of the proposed framework and assess the influence of various system parameter settings on accuracy.

Keyphrases: Lagrange coded computing, artificial noise, differential privacy, distributed learning

@booklet{EasyChair:12740,
  author    = {Yilei Xue and Jianhua Li and Jun Wu},
  title     = {DPLE: a Privacy-Enhanced and Straggler-Resilient Distributed Learning Framework for Smart Cloud},
  howpublished = {EasyChair Preprint 12740},
  year      = {EasyChair, 2024}}