Key Technologies For Privacy-Preserving Federated Learning Models For Indoor Localization In Edge Computing

With the increasing demand for location service,indoor location methods based on Received Signal Strength(RSS)fingerprint have been widely concerned and applied due to its mature infrastructure and easy implementation.However,RSS fingerprint indoor localization based on cloud architecture requires uploading a large amount of RSS data to the cloud server,which brings great challenges to the computing,communication and storage of the cloud server.In addition,these RSS data contains a large number of users’ sensitive information(such as location information,travel habits,etc.),and sending it directly to an untrusted cloud server will seriously infringe on users’ privacy and bring great trouble to users’ life and work.Besides,the performance of the Deep Learning(DL)model is closely related to the size of data sample and data richness.In order to provide users with high-performance indoor location services,a large amount of RSS data needs to be collected,but the collection and labeling of RSS data is time-consuming and labor-intensive.Therefore,how to provide users with highly reliable and strong real-time indoor location services under the premise of realizing fine-grained protection of user data privacy and model parameter privacy,reducing computing,communication,storage overhead,data labeling and collection tasks has become a major challenge for the rapid development of indoor location services.In response to this challenge,the main research works in this paper are as follows:(1)Aiming at the leakage of user privacy and model parameter privacy in the training and application process of DL-based indoor positioning model,and the difficulty of indoor positioning service based on cloud architecture to solve many challenges in the era of Internet of Everything,this paper proposes a differentially private federated learning model for fingerprinting indoor localization in edge computing(DP-FLoc EC),which uses FL technology to achieve local sub-model training,reduces data transmission delay while improving positioning accuracy,and uses Differential Privacy(DP)to solve user privacy leakage in model training and application stages question.Experimental results show that,compared with the centralized model based on cloud architecture,the method achieves higher localization accuracy and reduces communication overhead while providing provable privacy protection;compared with the distributed model based on FL architecture,the method provides more comprehensive privacy protection while achieving almost the same positioning accuracy and resource overhead.(2)In the process of collecting and marking RSS data based on cloud architecture,it will consume a lot of manpower,material resources,and financial resources.At the same time,these data contain a lot of user privacy.Direct exposure of these RSS data to the untrusted cloud will seriously infringe on user privacy.Aiming at this problem,this paper proposes a differentially private fingerprint fusion semi-supervised extreme learning machine for indoor localization in edge computing(Adp-FSELM).This method realizes the collection,fusion and privacy protection of RSS fingerprint data based on FL framework and DP,and uses a semi-supervised extreme learning machine(SELM)to implement indoor localization model training.The experimental results show that compared with the existing semi-supervised learning methods,this method can effectively reduce the amount of manual data annotation tasks while protecting the privacy of user data model parameters,and resisting Bayesian inference attacks.In addition,the method can provide users with real-time positioning services with less time overhead.(3)Given the high complexity and dynamics of the edge computing environment,the existing privacy budget allocation strategies are difficult to apply to the privacy protection of indoor positioning in the edge computing environment.By studying the privacy loss measurement and tracking method in the edge computing environment,this paper proposes a dynamic privacy budget allocation federated learning model for fingerprinting indoor localization in edge computing(ADP-FLoc EC)for the first time.The method uses Pearson correlation coefficient(PCC)and DP to collect and preprocess RSS data,and uses Rényi Differential Privacy(RDP)Dynamically to track the privacy loss during model training.The experimental results show that the proposed method can realize the fine distribution of the privacy budget in the indoor positioning model training process.While protecting the privacy of user data and model parameters,it obtains high indoor positioning model accuracy and low response delay.