| 摘要: |
| 随着物联网的发展，无人机辅助的边缘计算对于提高网络性能和数据处理能力至关重要。然而，在动态环境中，障碍物可能会导致无人机与用户之间的通信中断，间接影响用户的数据处理性能，对时间敏感的数据处理和用户隐私保护提出了挑战。针对该问题，提出了一种无人机辅助边缘计算的任务调度方案，旨在解决在涉及物理障碍的复杂条件下的时延成本优化和用户隐私问题。该方案融合局部差分隐私联邦学习(Local Differentially Private Federated Learning,LDP-FL)框架，通过协调用户任务调度、无人机轨迹和任务卸载率，在增强用户设备隐私保护的同时显著降低了系统的平均总时延成本。每个无人机-用户组都被分配了一个独立的深度强化学习代理，用于开发本地训练模型。然后，使用融合LDP-FL的加权聚合算法来处理和聚合它们的梯度，提高系统的时延性能和隐私安全性。与现有的将联邦学习与深度强化学习集成的算法相比，所提方案将时延成本降低了20.11%；此外，相对于用户设备的数量、随机任务的大小、无人机的计算能力和飞行持续时间，时延成本分别降低了25.46%、19.03%、14.59%和15.12%。 |
| 关键词: 无人机通信 任务调度；边缘计算 联邦学习 差分隐私 |
| DOI：10.20079/j.issn.1001-893x.240530002 |
|
| 基金项目:国家自然科学基金资助项目（62061003）;中国博士后科学基金项目(2022M722248)；中央高校基本科研业务费专项资金(J2023-027) |
|
| UAV-assisted Edge Computing Task Scheduling with Differential Privacy Federated Learning |
| LIU Jianhua,WANG Kexin,TU Xiaoguang,FAN Rong |
| (Institute of Electronic and Electrical Engineering,Civil Aviation Flight University of China,Guanghan 618300,China) |
| Abstract: |
| As the Internet of Things(IoT) expands,unmanned aerial vehicle(UAV)-assisted edge computing becomes crucial for enhancing network performance and data processing capabilities.However,in dynamic environments,obstacles may cause communication interruptions between UAV and users,indirectly affecting the data processing performance of user devices(UDs),posing challenges to time sensitive data processing and user privacy protection.In order to solve this problem,a task scheduling scheme for UAV-assisted edge computing is proposed,which aims to solve the problem of delay cost optimization and user privacy under complex conditions involving physical obstacles.By coordinating the scheduling of UDs,UAV trajectory,and task offloading rate based on local differential privacy federated learning(LDP-FL) framework,this scheme significantly reduces the average total delay cost of the system while enhancing UDs privacy protection.Each UAV-UDs group is assigned an independent deep reinforcement learning(DRL) agent that develops localized training models.Then,a weighted aggregation algorithm based on LDP-FL is used to process and aggregate their gradients to optimize these models and enhance system delay performance and privacy security.Compared with the existing algorithms integrating federated learning and DRL,the proposed scheme reduces the delay cost by 20.11%,and the delay cost is reduced by 25.46%,19.03%,14.59% and 15.12% respectively in term of the number of UDs,the size of random tasks,the computing power and flight duration of the UAV. |
| Key words: UAV communication task scheduling edge computing federated learning differential privacy |
