Adaptive Differential Privacy for Federated Learning via Quantization Noise
Year: 2026
Type: preprint
Abstract: Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without directly exchanging data. However, traditional differential privacy (DP) techniques, designed for centralized learning, often lead to substantial accuracy degradation in FL settings due to the inherent sensitivity of aggregated updates. This paper proposes a novel adaptive differential privacy (ADP) framework for FL that leverages quantization noise to mitigate this accuracy loss. Our approach dynamically adjusts the noise level based on the sensitivity of the local model updates, effectively reducing the overall noise while maintaining strong privacy guarantees. We demonstrate through theoretical analysis and simulations that our ADP framework achieves a superior trade-off between privacy and accuracy compared to standard DP approaches in FL. This work represents a significant advancement in the field by combining quantization and adaptive mechanisms for enhanced accuracy in federated learning.
Author: Zhang, Jincheng
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7210267836