Research On Privacy-preserving Federated Recommendation Based On Machine Learnin

With the rapid development of the Internet,users are overwhelmed by the growing amount of data,which brings the phenomenon of information overload.Recommendation systems play an important role in alleviating information overload,and their analytical use of user data,while providing efficient recommendation services,also poses the risk of user privacy leakage.The emergence of new machine learning paradigms such as federation learning and graph neural networks provides new ideas and solutions for the implementation of recommendation systems,which can realize the effective utilization of user data by recommendation system models while protecting data security.On this basis,two recommendation schemes are proposed in this thesis,as follows:(1)A federated recommendation scheme based on personalized user grouping is proposed.Considering the timeliness of users receiving recommendation services,a group of users with the same interests is used to assist in training the recommendation model,which achieves fast and personalized training of the recommendation model and effectively solves the problems of high model training communication consumption and insufficient model personalization caused by the heterogeneity of user data and data sparsity of the federated recommendation system.Specifically,users are grouped based on user similarity and device data volume to collaboratively train models and improve model personalization;in model aggregation,different aggregation weights are assigned based on data and performance status among different clients to reduce communication burden and accelerate model training at the same time.(2)A federated recommendation scheme based on graph neural network is proposed,which can effectively improve the recommendation performance while achieving privacy protection.This scheme applies graph neural networks to federated recommendation system and uses high-order information to build recommendation models,effectively solving the problem of poor performance of recommendation systems under privacypreserving scenarios.Specifically,server extracts user preference features from the data protected by differential privacy,groups users with similar preferences and extracts highorder information.The high-order information is introduced into the client to ensure the adequacy of local data,make full use of the device performance by increasing the number of local training rounds,and reduce the communication consumption while making the model training more accurate.In addition,the privacy protection scheme is designed to achieve the protection of users’ private data.Finally,the two schemes are experimentally validated on two real datasets.The results show that the two schemes proposed in this thesis can achieve better recommendation accuracy while protecting user privacy.