A HYBRID SECURE AGGREGATION AND DIFFERENTIAL PRIVACY FRAMEWORK FOR COMMUNICATION-EFFICIENT BIG DATA ANALYTICS
JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES · 24.06.2026 · 21 (6)
DOI: 10.26782/jmcms.2026.06.00003 ↗Kurz gefasst
Fachartikel in der Zeitschrift JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES, online erschienen am 21.06.2021. Verfasst von Subhajit Roy und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Federated Learning (FL) is a type of distributed learning where several clients (clusters) train a machine learning model without sharing the raw data directly. Nevertheless, there are still three significant issues for practical FL systems: privacy leakage through sharing the model update, heavy communication burden, and unstable learning performance when the data distribution is not IID. To enhance the protection of privacy and communication efficiency in distributed big data analytics, the authors introduce a Privacy-Preserving Federated Learning (PP-FL) framework combining Differential Privacy (DP), Secure Aggregation (SA), and Adaptive Gradient Compression (AGC). Differential Privacy implies that calibrated Gaussian noise is added to local updates, and Secure Aggregation ensures that the central entity cannot see individual updates from the clients. Adaptive Gradient Compression cuts down on communicating the most important components of a gradient. The proposed framework is tested with the high-resolution MNIST dataset and CIFAR-10 dataset in non-IID federated settings. The experimental results demonstrate that PP-FL has a significant boost in reducing communication cost compared to standard FedAvg and DP-FedAvg. The results also illustrate a clear trade-off between privacy and utility, as the addition of noise (differential privacy) or departure from an IID data distribution or extreme compression can negatively impact classification accuracy. This behaviour is explained by means of revised studies that consist of a component-wise interpretation, observation of the gradient norm, and evaluation of the compression ratio. The overall results suggest that the proposed PP-FL system can be employed to implement communication-efficient and privacy-preserving federated learning, and that careful tuning of noise parameters and compression parameters is essential to ensure the stability of learning.
Zitieren
APA
Roy, S., Chakraborty, R., & Chowdhury, T. (2026). A HYBRID SECURE AGGREGATION AND DIFFERENTIAL PRIVACY FRAMEWORK FOR COMMUNICATION-EFFICIENT BIG DATA ANALYTICS. JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES, 21(6). https://doi.org/10.26782/jmcms.2026.06.00003
Vancouver
Roy S, Chakraborty R, Chowdhury T. A HYBRID SECURE AGGREGATION AND DIFFERENTIAL PRIVACY FRAMEWORK FOR COMMUNICATION-EFFICIENT BIG DATA ANALYTICS. JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES. 2026;21(6). doi:10.26782/jmcms.2026.06.00003
BibTeX
@article{Roy2026,
author = {Roy, Subhajit and Chakraborty, Rupak and Chowdhury, Tapan},
title = {A HYBRID SECURE AGGREGATION AND DIFFERENTIAL PRIVACY FRAMEWORK FOR COMMUNICATION-EFFICIENT BIG DATA ANALYTICS},
journal = {JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES},
year = {2026},
volume = {21},
number = {6},
doi = {10.26782/jmcms.2026.06.00003}
}
EUVIMED