Privacy-preserving Recommendation System Based On Federated Learning

With the rapid development of the internet,the amount of data had a significant increase in the past few years,causing the problem of information overload.Recommendation System(RS)came into being to solve the problem of information overload.However,user privacy may be leaked in RS.Although privacy-preserving methods based on data perturbation and Homomorphic Encryption(HE)can protect users’ privacy to a certain extent,they still cannot meet the requirements of privacy preservation,accuracy,and time efficiency simultaneously.Federated Learning(FL)provides a new solution for the privacy-preserving RS.Nevertheless,the existing federated RS still has the risk of leaking user review information and inaccurate recommendation results.To solve the above problems,this thesis proposes a Partition-based Federated Hybrid Collaborative Filtering Model(PFHCF).The main research contents are as follows:(1)Aiming at user privacy and security issues,this thesis improves the training framework of FL and proposes a partition-based mechanism to divide users and items into multiple groups.Each user group is responsible for training all item groups in turn.Under this training mecha-nism,when users want to obtain item information,no interacted item ID will be upload to the server as a parameter,therefore,neither the third party nor the server can know which items the user has commented on.Thus this mechanism can ensure the security of user information.(2)Aiming at the problem of poor recommendation,this thesis proposes a pre-training mechanism based on Denoising Auto Encoder(DAE)to fuse auxiliary information of items.Since the auxiliary information of items in the FL is stored on the server,DAE is used to pre-train the auxiliary information of items.Then the recommendation model serves the output of the encoder as the initial parameters of the embedding layer.Consequently,the model parameters are fine-tuned in the subsequent training,which helps improve the recommendation accuracy without increasing the communication cost between the client and the server.Besides,this thesis proposes a novel Hybrid Neural Collaborative Filtering Model(HNCF).Because the client-side user information in FL is stored locally and easy to collect,the user features are added to the input layer of HNCF.Element-wise product and multi-layer perceptron are integrated into the interaction layer of the model.The model can learn abundant interactive information,therefore improve the recommendation effect.To verify the effectiveness of PHFCF,this thesis conducted comparative experiments on real-world datasets.The experimental results show that the PHFCF model can effectively pre-vent data leakage and obtain better performance.