Privacy‐Preserving Federated Learning: Adaptive Quantization and Layer‐Specific Differential Privacy With Client Dropout Mitigation
Applied Computational Intelligence and Soft Computing · 2026-01 · 2026 (1)
DOI: 10.1155/acis/1125376 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 30.05.2026. Verfasst von Ahmad G. Atia und 3 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
This paper presents an enhanced federated learning architecture that addresses key challenges in distributed machine learning through three novel contributions: (1) a theoretically grounded sensitivity analysis for differential privacy that formally derives the Δ parameter for deep learning models, (2) an adaptive quantization scheme that dynamically adjusts to model architecture and layer importance, and (3) a client dropout mitigation strategy using weighted aggregation and historical model averaging. We evaluate our framework on four benchmark datasets, demonstrating superior performance in a unified comparative analysis over FATE and PySyft frameworks. The system integrates AES, RSA, and TLS encryption with differential privacy in a novel framework that maintains model accuracy alongside strong privacy guarantees. We provide a comprehensive theoretical analysis of the sensitivity parameter Δ for convolutional neural networks, demonstrating its dependence on layer depth and gradient norms. Our adaptive quantization approach shows 15%–20% higher compression efficiency compared to uniform quantization while maintaining model accuracy. Experimental results on MNIST, CIFAR‐10, CIFAR‐100, and Fashion‐MNIST demonstrate the system’s effectiveness, achieving accuracies of 98.5%, 85.3%, 65.5%, and 92.1%, respectively, while maintaining ϵ‐differential privacy with ϵ values between 0.3 and 1.0, with formal (ε, δ)‐differential privacy guarantees tracked via the moments accountant method. The proposed client dropout handling method reduces accuracy degradation by up to 40% compared to standard federated averaging under 30% client dropout rates. Comparative analysis with FATE and PySyft shows our approach achieves 16−17% better communication efficiency while maintaining comparable privacy guarantees.
Zitieren
APA
Atia, A. G., Karawia, A. A., Fouda, Y. M., & El-Gayar, M. M. (2026). Privacy‐Preserving Federated Learning: Adaptive Quantization and Layer‐Specific Differential Privacy With Client Dropout Mitigation. Applied Computational Intelligence and Soft Computing, 2026(1). https://doi.org/10.1155/acis/1125376
Vancouver
Atia AG, Karawia AA, Fouda YM, El-Gayar MM. Privacy‐Preserving Federated Learning: Adaptive Quantization and Layer‐Specific Differential Privacy With Client Dropout Mitigation. Applied Computational Intelligence and Soft Computing. 2026;2026(1). doi:10.1155/acis/1125376
BibTeX
@article{Atia2026,
author = {Atia, Ahmad G. and Karawia, Abdelrahman A. and Fouda, Yasser M. and El-Gayar, Mostafa M.},
title = {Privacy‐Preserving Federated Learning: Adaptive Quantization and Layer‐Specific Differential Privacy With Client Dropout Mitigation},
journal = {Applied Computational Intelligence and Soft Computing},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/acis/1125376}
}
EUVIMED