Due to the limited storage and computing power, edge devices at the network edge cannot train deep learning models locally. Traditional deep learning training requires users to upload a local dataset to a cloud center, and trains the data using massive computation resources of the cloud center. However, it results in two bad effects: uploading a local dataset to a centralized cloud center controlled by a third party leaves user data privacy at risk; and uploading multimedia data will consume huge bandwidth resources of mobile users and storage resources of the cloud center, resulting in low scalability in term of the number of edge devices. To deal with these two problems, we propose an edge-enabled distributed deep learning platform by dividing a general deep learning training network into a front and back subnetwork. Specifically, the front subnetwork consisting of several layers is deployed close to input data and is trained separately at each edge device using the local dataset, and the outputs of all front subnetworks are sent to the back subnetwork for later training at a cloud center; while the back subnetwork is deployed at the cloud center, and its output is sent to each front subnetwork. As no original dataset is transferred from edge devices to the cloud center, the platform can protect data privacy and has high scalability. Above that, another two measures are taken to ensure data privacy: asymmetric encryption technology is adopted to guarantee the safety and integrality of the transferred parameters between edge servers and the cloud center; and blockchain technology is used to monitor the actions of the stakeholders in this platform and thereby ensure trust among the stakeholders. Experimental results show the validation of the proposed method.

中文翻译：

---

用于 5G 隐私保护的边缘分布式深度学习


由于存储和计算能力有限，网络边缘的边缘设备无法在本地训练深度学习模型。传统的深度学习训练需要用户将本地数据集上传到云中心，利用云中心的海量计算资源来训练数据。然而，这会带来两个不好的影响：将本地数据集上传到第三方控制的集中式云中心，使用户数据隐私面临风险；而上传多媒体数据会消耗移动用户巨大的带宽资源和云中心的存储资源，导致边缘设备数量的可扩展性较低。为了解决这两个问题，我们通过将通用深度学习训练网络分为前子网和后子网，提出了一种边缘支持的分布式深度学习平台。具体来说，由多层组成的前子网络部署在靠近输入数据的位置，并使用本地数据集在每个边缘设备上单独进行训练，所有前子网络的输出都发送到后子网络，以便稍后在云中心进行训练；后端子网部署在云中心，其输出发送到各个前端子网。由于没有原始数据集从边缘设备传输到云中心，因此该平台可以保护数据隐私并具有较高的可扩展性。除此之外，还采取了两项措施来确保数据隐私：采用非对称加密技术，保证边缘服务器与云中心之间传输参数的安全性和完整性；区块链技术用于监控该平台中利益相关者的行为，从而确保利益相关者之间的信任。实验结果验证了所提方法的有效性。