边缘计算下指纹室内定位差分私有联邦学习模型

A Differentially Private Federated Learning Model for Fingerprinting Indoor Localization in Edge Computing

随着人们对位置服务需求的日益增长,基于接收信号强度(received signal strength,RSS)指纹的室内定位技术因具有其成熟的基础设施和易于实现等优势而受到广泛关注.深度学习(deep learning,DL)强大的特征抽取和自动分类能力使其成为基于RSS指纹室内定位的一个非常有吸引力的方案.但是,这种方案需要使用大量的RSS指纹数据并借助云计算对DL模型进行重复训练.由于RSS数据包含了用户的个人敏感信息,直接将这些数据发送到不可信的云端进行处理,会造成严重的用户隐私侵犯和数据传输延迟.针对以上挑战,提出了一种边缘计算下指纹室内定位差分私有联邦学习模型.该模型构建了边缘计算下的联邦学习协议并设计了一个基于卷积神经网络(convolutional neural network,CNN)的轻量级室内定位模型,不再需要将大量的RSS数据上传到云端后进行模型训练,在提高定位精度的同时减少数据传输延迟;然后,利用差分隐私技术解决了离线训练阶段和在线定位阶段的用户隐私泄露问题.在多个真实数据集上的实验结果和安全性分析表明,与基于云架构的集中式模型相比,该机制在提供可证明的隐私保护情况下取得了较高的定位精度、减少了通信开销;与基于联邦学习架构的分布式模型相比,该机制在取得几乎相同定位精度和资源开销的情况下,提供了更全面的隐私保护.

With the increasing demand for location-based services,fingerprinting indoor localization based on received signal strength(RSS)has attracted widespread attention due to its well-established infrastructures and easy implementation.Deep learning(DL)has become a very attractive solution for RSS-based fingerprinting indoor localization because of its powerful capabilities of feature extraction and automatic classification.However,these solutions require repeated training of DL model with a large amount of RSS fingerprinting data via cloud computing.Since the RSS data contain users personal sensitive information,it may cause serious users privacy violations and data transmission delays when sending these RSS data directly to the untrusted cloud for processing.To address these challenges,a differentially private federated learning model for fingerprinting indoor localization in edge computing(DP-FLocEC)is proposed in this paper,which builds an edge computing enabled federated learning protocol and a convolutional neural network(CNN)based lightweight indoor localization model.The DP-FLocEC does no longer need to upload a large amount of RSS data to the cloud for model training,which improves localization accuracy while reducing data transmission delay.Then,we employ differential privacy technology to solve the problem of user privacy leakage in the offline training phase and the online localization phase of indoor localization.Security analysis and experimental results on multiple real datasets show that,compared with the centralized model based on cloud architecture,the DP-FLocEC obtains higher localization accuracy and reduces communication loss while providing provable privacy-preserving;compared with the distributed model based on the federated learning architecture,our DP-FLocEC provides more comprehensive privacy-preserving for users with almost the same localization accuracy and resource overhead.