Moderation is the Best Policy: Dynamic Defense Against Gradient-Based Data Reconstruction Attacks in Federated Learning
- 9 February 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (3) , 6230-6245
Abstract
Federated learning (FL) is a privacy-preserving distributed machine learning framework. However, recent studies have shown that implementing gradient-based data reconstruction attacks (DRA) can still lead to the leakage of user privacy through frequently uploaded model parameters in FL. Existing works leverage differential privacy (DP) to prevent privacy leakage, but the lack of effective scheduling of the privacy budget results in significant accuracy loss in the trained models. In this paper, we propose a novel dynamic privacy preserving federated learning framework, named NDPP-FL, capable of delivering robust defenses against DRA while significantly mitigating performance loss. Our key insight is to regard the privacy budget as a non-replenishable resource and dynamically schedule it based on privacy leakage risks to provide self-adaptive privacy protection for clients across varying communication rounds. Specifically, based on the amount of information between the local dataset and the transmitted parameters, we first design a parameter channel information leakage model. Then, during each update iteration, we introduce saliency perturbations based on the Hessian matrix to enhance defensive capabilities. Meanwhile, to improve the performance of NDPP-FL, sample-adaptive clipping and decaying noise perturbations are adopted in the construction. Furthermore, extensive experiments demonstrate that our framework performs excellently in terms of model accuracy and resilience against DRA.Keywords
Funding Information
- Young Talent Fund of Xi’an Association for Science and Technology (0959202513203)
- the RC (EP/Y027787/1)
- National Natural Science Foundation of China (62272391)
- National Natural Science Foundation of China (62402391)
- National Natural Science Foundation of China (62402388)
- Fundamental Research Funds for the Central Universities (D5000250275)
- Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China (JYB2025XDXM116)
- Shenzhen Science and Technology Program (CJGJZD20220517142005013)
This publication has 39 references indexed in Scilit:
- Make Landscape Flatter in Differentially Private Federated Learning2023
- Auditing Privacy Defenses in Federated Learning via Generative Gradient Leakage2022
- Deep Leakage from Gradients2020
- A Framework for Evaluating Client Privacy Leakages in Federated Learning2020
- The Science of Quantitative Information Flow2020
- Rényi Differential Privacy2017
- Differential Privacy2017
- Concentrated Differential Privacy: Simplifications, Extensions, and Lower Bounds2016
- Deep Residual Learning for Image Recognition2016
- Gradient-based learning applied to document recognitionProceedings of the IEEE, 1998