Title: Security-enhanced machine learning framework based on PATE

Authors: Xian Guo; Kai Zheng; Yongbo Jiang; Jing Wang; Junli Fang

Addresses: School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China ' School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China ' School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China ' School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China ' School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China

Abstract: Privacy aggregated teacher ensembles (PATE) is a general machine learning framework that provides privacy-preserving for training data. However, this framework faces security risks in the distributed learning environment. Firstly, the involvement of illicit nodes in communication may lead to aggregation result inaccuracies. Secondly, the semi-honest aggregator and teacher nodes could potentially result in privacy leaks of other teacher nodes. Thirdly, the aggregation results are influenced by each teacher, and there may be poisoning attacks during the aggregation process. Fourthly, malicious aggregator may tamper with the information sent to student nodes or attempt to access relevant information about student node training labels. To address the above issues, we propose a machine learning framework with stronger security and privacy in a distributed learning environment based on principal component analysis and secures multi-party computing. The framework is subjected to security analysis and experimental validation. The security analysis establishes the framework's robustness and privacy-preserving characteristics, while experimental validation demonstrates its practical viability.

Keywords: distributed learning; private aggregation of teacher ensemble; privacy-preserving; machine learning.

DOI: 10.1504/IJICS.2025.145126

International Journal of Information and Computer Security, 2025 Vol.26 No.1/2, pp.109 - 146

Received: 23 Sep 2023
Accepted: 14 Apr 2024
Published online: 19 Mar 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article