| 摘 要: 为了解决联邦学习中的边缘异构性和隐私泄露的问题,提出了一种基于差分隐私的半异步联邦学习(FedSA-DP)机制。在通信优化层面,设计动态加权半异步聚合机制,通过松弛同步约束允许部分节点延迟更新,结合设备计算能力与信道状态的自适应权重分配,比传统FedAvg算法有效缩短整体训练时间74.8%。在隐私保护层面,创新性地构建多阶段高斯噪声注入方案,将差分隐私(DP)嵌入本地模型训练与参数传输全过程,在(ε,δ)-隐私保障下非独立同分布(Non-IID)数据测试准确率较 FedAvg提升了26.8%。实验结果表明,该方法在 MNIST 和CIFAR-10数据集上达到相同准确率的同时,相比传统算法缩短训练耗时至多74.8%,为边缘计算场景下隐私与效率的协同优化提供了新的技术路径。 |
| 关键词: 半异步联邦学习 差分隐私 数据异构 边缘计算 |
|
中图分类号: TP301
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(62371256） |
|
| Edge Computing Method for Semi-Asynchronous Federated Learning with Differential Privacy |
|
JIANG Yunyu, JIANG Zhefan, WANG Dazhi, WU Yuanhao, XIAO Jian
|
(College of Integrated Circuit Science and Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210023, China)
1255003271@qq.com; 1754054595@qq.com; 1538566617@qq.com; 1915583851@qq.com; xiaoj@njupt.edu.cn
|
| Abstract: To address the problems of edge heterogeneity and privacy leakage in federated learning, a Sem-i Asynchronous Federated learning based on Differential Privacy ( FedSA-DP) mechanism is proposed. At the communication optimization level, a dynamic weighted sem-i asynchronous aggregation mechanism is designed. By relaxing synchronization constraints to allow partial node delayed updates and combining adaptive weight allocation based on device computational capability and channel state, this mechanism effectively reduces the total training time by 74.8% compared to the traditional FedAvg algorithm. At the privacy protection level, an innovative mult-i stage Gaussian noise injection scheme is constructed, embedding Differential Privacy (DP) throughout the entire process of local model training and parameter transmission. This achieves a 26.8% improvement in test accuracy for Non-IID (Non-Independent and Identically Distributed) data under the (ε,δ)-privacy guarantee compared to FedAvg. Experiment results show that while achieving the same accuracy on the MNIST and CIFAR-10 datasets, this method reduces training time by up to 74.8% compared to traditional algorithms, providing a new technical pathway for the collaborative optimization of privacy and efficiency in edge computing scenarios. |
| Keywords: Sem-i Asynchronous Federated Learning Differential Privacy data heterogeneity edge computing |
