CEEP-FL: A comprehensive approach for communication efficiency and enhanced privacy in federated learning
CEEP-FL: A comprehensive approach for communication efficiency and enhanced privacy in federated learning
复制标题
DOI：
10.1016/j.asoc.2021.107235
复制
发表时间：
2021-03-06
影响因子：
8.7
通讯作者：
Aslam, Muhammad
中科院分区：
文献类型：
--
作者：
Asad, Muhammad;Moustafa, Ahmed;Aslam, Muhammad
联邦学习（FL）是一种新兴的技术，用于在隐私约束下在分布式数据上协作训练机器学习模型。然而，最近的研究表明，FL显着消耗大量的通信资源，在全球模型更新。此外，参与者的私人数据也可以通过利用共享参数上传到中央云服务器时的本地梯度更新，这阻碍了FL被广泛实施。为了应对这些挑战，在本文中，我们提出了一种新的全面FL方法，即通信效率和增强隐私（CEEP-FL）。特别是，所提出的方法同时旨在：（1）最小化通信成本，（2）保护数据不被泄露，以及（3）最大化全局学习精度。为了最小化通信成本，我们首先在每个局部梯度更新上应用一种新的过滤机制，并仅上传重要的梯度。然后，我们应用基于同态密码系统（NIZKP-HC）的非交互式零知识证明，以保护这些局部梯度更新，同时保持网络的鲁棒性。最后，我们使用分布式选择性随机梯度下降（DSSGD）优化，以最大限度地减少计算成本和最大限度地提高全局学习精度。在常用外语数据集上的实验结果表明，CEEP-FL方法明显优于现有方法。(c)2021爱思唯尔有限公司版权所有。
Federated Learning (FL) is an emerging technique for collaboratively training machine learning models on distributed data under privacy constraints. However, recent studies have shown that FL significantly consumes plenty of communication resources during the global model update. In addition, participants' private data can also be compromised by exploiting the shared parameters when uploading the local gradient updates to the central cloud server, which hinders FL to be implemented widely. To address these challenges, in this paper, we propose a novel comprehensive FL approach, namely, Communication Efficient and Enhanced Privacy (CEEP-FL). In particular, the proposed approach simultaneously aims to; (1) minimize the communication cost, (2) protect data from being compromised, and (3) maximize the global learning accuracy. To minimize the communication cost, we first apply a novel filtering mechanism on each local gradient update and upload only the important gradients. Then, we apply Non-Interactive Zero-Knowledge Proofs based Homomorphic-Cryptosystem (NIZKP-HC) in order to protect those local gradient updates while maintaining robustness in the network. Finally, we use Distributed Selective Stochastic Gradient Descent (DSSGD) optimization to minimize the computational cost and maximize the global learning accuracy. The experimental results on commonly used FL datasets demonstrate that CEEP-FL distinctively outperforms the existing approaches. (c) 2021 Elsevier B.V. All rights reserved.