Research On Privacy-preserving Deep Learning Recommendation Algorithm Based On Federated Learnin

Recommender systems are an effective way to handle the problem of information overload.However,collecting massive user data to train the model poses challenges to data security and user privacy protection.Federated learning can collaborate on training the recommender model without revealing user privacy by using user data that does not leave their local device,thereby improving recommendation accuracy.With the rapid development of deep learning,deep learning-based recommendation algorithms can efficiently handle massive training samples and integrate multiple types of additional information,alleviating the inherent data sparsity and cold-start problems in recommender systems.In recent years,research on federated learning-based recommendation algorithms has made some progress,but there are still some issues that need to be addressed.The main research contents of this thesis are as follows:(1)Existing research on federated learning-based recommendation algorithms mainly focuses on combining traditional collaborative filtering matrix factorization algorithms with federated learning frameworks.These approaches only utilize user-item interactions while ignoring rich feature information.There is also relatively little research on the combination of deep learning-based recommendation algorithms and federated learning.This thesis attempts to explore federated recommendation algorithms that combine deep learning techniques while also utilizing more feature information.Specifically,this thesis proposes a federated deep learning recommendation algorithm called Fed Deep FM,which combines the model parameter transmission-based federated learning framework with the deep learning-based recommendation algorithm Deep FM.It can not only incorporate general features of users and items into the training but also leverage the powerful ability of deep learning to fully explore high-order interactions between features,thereby providing high-quality recommendations.Fed Deep FM uploads the model parameters of each client instead of raw data or gradients for model updating.This ensures that the server can train a globally generalizable model using the privacy data held by each client without directly collecting user interaction information.Moreover,according to preliminary experiments,compared with the commonly used gradient uploading method in federated recommendation algorithms,the method of uploading model parameters does not reveal the feature composition of the local dataset directly due to the issue of "gradient being 0 for feature values that do not appear in the training set." Experimental results show that Fed Deep FM can provide high-quality recommendations by combining general features of users and items and deep learning technology without collecting local user data.(2)The problem with using only the model training parameters rather than the raw data to achieve privacy protection in federated learning has been proven to be vulnerable to various new targeted attacks.This thesis attempts to combine existing or propose novel privacy protection methods within the implemented federated recommendation algorithm framework to further enhance its privacy protection level.Specifically,this thesis proposes a privacy protection method based on perturbed data,called pseudo interaction padding,to provide users with stronger privacy protection when facing inference attacks initiated by honest but curious entities in the federated recommendation scenario.This method generates and expands the client’s training data set through a random process to mask the real user data,which is jointly used for training.Theoretical analysis combined with experimental results shows that Fed Deep FM,which combines the privacy protection method based on perturbed data,pseudo interaction padding,can enhance user privacy while ensuring high-quality recommendations.