A Privacy-Preserving Federated Learning for Multiparty Data Sharing in Social IoTs
A Privacy-Preserving Federated Learning for Multiparty Data Sharing in Social IoTs
复制标题
社交物联网中多方数据共享的隐私保护联合学习
DOI：
10.1109/tnse.2021.3074185
复制
发表时间：
2021-07-01
影响因子：
6.6
通讯作者：
Cheng, Xiaochun
中科院分区：
文献类型：
--
作者：
Yin, Lihua;Feng, Jiyuan;Cheng, Xiaochun
随着5G和移动的计算的快速发展，社交计算和社交物联网（IoT）中的深度学习服务在过去几年中丰富了我们的生活。具备计算能力的移动的设备和物联网设备可以随时随地加入社交计算。联合学习允许充分使用分散的培训设备，而无需原始数据，从而为进一步打破数据孤岛和提供更精确的服务提供便利。然而，各种攻击表明，目前联邦学习的培训过程仍然受到数据和内容层面的披露的威胁。本文提出了一种新的混合隐私保护方法，用于联邦学习以应对上述挑战。首先，我们采用了一种先进的函数加密算法，不仅保护了每个客户端上传的数据的特性，而且还保护了加权求和过程中每个参与者的权重。噪声机制通过设计局部贝叶斯差分隐私机制，有效地提高了对不同分布数据集的适应性。此外，在联邦学习训练中，我们还使用稀疏差分梯度来提高传输和存储效率。实验表明，当我们使用稀疏差分梯度来提高传输效率时，模型的准确性最多只下降了3%。
As 5G and mobile computing are growing rapidly, deep learning services in the Social Computing and Social Internet of Things (IoT) have enriched our lives over the past few years. Mobile devices and IoT devices with computing capabilities can join social computing anytime and anywhere. Federated learning allows for the full use of decentralized training devices without the need for raw data, providing convenience in breaking data silos further and delivering more precise services. However, the various attacks illustrate that the current training process of federal learning is still threatened by disclosures at both the data and content levels. In this paper, we propose a new hybrid privacy-preserving method for federal learning to meet the challenges above. First, we employ an advanced function encryption algorithm that not only protects the characteristics of the data uploaded by each client, but also protects the weight of each participant in the weighted summation procedure. By designing local Bayesian differential privacy, the noise mechanism can effectively improve the adaptability of different distributed data sets. In addition, we also use Sparse Differential Gradient to improve the transmission and storage efficiency in federal learning training. Experiments show that when we use the sparse differential gradient to improve the transmission efficiency, the accuracy of the model is only dropped by 3% at most.