Time-Varying Noise Perturbation and Power Control for Differential-Privacy-Preserving Wireless Federated Learning
Time-Varying Noise Perturbation and Power Control for Differential-Privacy-Preserving Wireless Federated Learning
复制标题
DOI：
10.1109/ieeeconf59524.2023.10476780
复制
发表时间：
2023-10
期刊：
影响因子：
--
通讯作者：
Dang Qua Nguyen;Taejoon Kim
中科院分区：
文献类型：
--
作者：
Dang Qua Nguyen;Taejoon Kim
无线联合学习（FL）是一个框架，它使服务器能够通过无线信道与分布式用户协作训练学习模型，而无需共享用户的训练数据。然而，由于工程反演攻击，用户的数据隐私泄漏在训练过程中是一个至关重要的问题。差分隐私（DP）技术通常用于处理这个问题。然而，这种方法可能会导致学习效用的下降。本文提出了一种隐私保护差分隐私FL算法，应用时变噪声方差扰动。利用现有的无线信道噪声，我们联合设计DP噪声方差和用户的发射功率，以解决隐私和学习效用的权衡在无线FL。此外，FL迭代的次数进行了优化，通过最小化上界的学习误差。我们进行模拟，以证明我们的方法在DP保证和学习效用方面的有效性。
Wireless Federated Learning (FL) is a framework that enables a server to collaboratively train a learning model with distributed users via wireless channels without sharing users' training data. However, due to the engineering-inversion attack, users' data privacy leakage during the training process is a crucial concern. Differential privacy (DP) techniques are commonly applied to deal with this issue. This approach, however, can cause degradation in the learning utility. This paper proposes a privacy-preserving differentially private FL algorithm that applies time-varying noise variance perturbation. Taking advantage of the existing wireless channel noises, we jointly design DP noise variances and users' transmit power to address privacy and learning utility tradeoffs in the wireless FL. In addition, the number of FL iterations is optimized by minimizing the upper bound on the learning error. We conduct simulations to demonstrate the effectiveness of our approach in terms of DP guarantee and learning utility.