Research On Crowd Sensing Data Security Method Based On Federated Learning

The role of analysis,reasoning and decision making based on artificial intelligence has become indispensable and more and more important in the process of completing large-scale and fine-grained perceptual tasks in the new generation of crowd sensing systems.Mobile Crowd Sensing is an open system that relies on mobile intelligent terminal devices to participate in data perception computing,which is easy to cause user privacy data leakage.Federated Learning(FL)is a new distributed machine learning method that enables a group of devices to train a shared artificial intelligence model together.In the process of model training,user data is saved locally and only model parameters need to be uploaded,which greatly improves the security of user data.However,there may be malicious users in the Federated Learning system and the process of data interaction will lead to unnecessary data leakage.In order to effectively protect data security in Crowdsensing applications,this paper studies crowd sensing data security methods based on Federated Learning.The main work is as follows:(1)Structure a federated learning model : Based on the application scenarios and characteristics of Crowdsensing,in order to effectively protect the data security of crowd sensing,this paper introduces the Federated Learning model into the crowd sensing system,so that users’ data only needs to be saved locally to prevent the risk of user data leakage.Therefore,according to the FL principle,a Federated Learning model suitable for crowd sensing is designed using Tensor Flow framework.(2)Structure an Efficient Federated Learning Framework for Group Intelligence Perception:The training process of federated learning model requires repeated and frequent model data exchange between the server and the client,which will cause a large overhead.This paper proposes an efficient method for selecting training equipment of participation model,The improved K-Means clustering method was used to cluster all the devices involved in the Sensing.On this basis,the cluster heads were randomly selected as the representative devices to participate in the subsequent model training.The experimental results show that compared with the traditional method,the system overhead is reduced by about 20% on the basis of small accuracy loss(about 2% accuracy loss).(3)Crowd Sensing Data Security Method based on Federated Learning:In view of the adverse impact on system model accuracy and data security caused by malicious users and potential security risks of model data interaction in Federated Learning,in order to build a secure Federated Learning framework for crowd sensing data security protection,this paper proposes a legal user screening method based on reputation scores.By calculating the gradient similarity between different models,the contribution degree of the user model to the total server model is evaluated,so as to judge whether the user is reliable.The ability of resisting malicious user attack is verified by different attack experiments.At the same time,through comparative experiments,this paper makes it clear that appropriate encryption methods can be selected according to the size of data in the application of crowd sensing to effectively protect the security of data in transmission.