摘要：

联邦学习技术的飞速发展促进不同终端用户数据协同训练梯度模型，其显著特征是训练数据集不离开本地设备，只有梯度模型在本地进行更新并共享，使边缘服务器生成全局梯度模型。然而，本地设备间的异构性会影响训练性能，且共享梯度模型更新具有隐私泄密与恶意篡改威胁。提出云-边融合的可验证隐私保护跨域联邦学习方案。在方案中，终端用户利用单掩码盲化技术保护数据隐私，利用基于向量内积的签名算法产生梯度模型的签名，边缘服务器通过盲化技术聚合隐私数据并产生去盲化聚合签名，确保全局梯度模型更新与共享过程的不可篡改性。采用多区域权重转发技术解决异构网络中设备计算资源与通信开销受限的问题。实验结果表明，该方案能够安全高效地部署在异构网络中，并在MNIST、SVHN、CIFAR-10和CIFAR-100 4个基准数据集上进行系统实验仿真，与经典联邦学习方案相比，在精度相当的情况下，本文方案梯度模型收敛速度平均提高了21.6%。

关键词: 联邦学习, 全局梯度模型, 数据隐私, 可验证隐私保护, 跨域训练

Abstract:

The rapid development of Federated Learning(FL) technology promotes collaborative training of gradient models using data from different end users. Its notable feature is that the training dataset does not leave the local device, and only gradient model updates are locally computed and shared, enabling edge servers to generate global gradient models. However, the heterogeneity between local devices can affect training performance, and shared gradient model updates pose privacy breaches and malicious tampering threats. This study proposes a verifiable privacy-preserving cross-domain FL scheme based on cloud-edge fusion. In the scheme, end users use single mask blinding technology to protect data privacy, vector inner product based signature algorithms to generate signatures for gradient models, and edge servers aggregate private data through blinding technology to generate deblinded aggregated signatures. This ensures the global gradient model is updated and the sharing process is tamper proof. It adopts multi-region weight forwarding technology to address the problem of limited computing resources and communication costs of devices in heterogeneous networks. The experimental results demonstrate that the proposed scheme can be safely and efficiently deployed in heterogeneous networks, and system experiments and simulations are performed on four benchmark datasets: MNIST, SVHN, CIFAR-10, and CIFAR-100. Compared with the classical federated learning scheme, the gradient model convergence speed of our scheme is improved by an average of 21.6% with comparable accuracy.

Key words: Federated Learning(FL), global gradient model, data privacy, verifiable privacy-preserving, cross-domain training