Research And Application Of The Recommendation System Based On Differential Privacy

The recommendation system can help users filter out items that are more in line with their personal wishes from the massive data.At the same time,the recommendation results will gradually improve as users continue to choose and reject.However,a large amount of user private information will also be generated in this process,and this private information will inevitably face some privacy risks during the use of the recommendation system.Once user privacy is not guaranteed and some negative cases appear,it will shake the user's trust in the system,which will result in reducing user activity or providing untrue data,which will make the recommendation effect worse,and will lead to a large-scale loss of users.The project will be unsustainable.This article focuses on the problem of privacy protection in recommender systems,mainly from the following three aspects.Firstly,in the existing Personal Rank process,a bipartite graph is obtained based on the operations of different users on items,and then the transition matrix iteration is obtained from the bipartite graph,but the privacy of the recommended target user cannot be guaranteed in this process.Therefore,using the differential privacy mechanism,Laplace noise is added to the user's score,and the weight of the walking is recalculated.After the iterative calculation,you can directly sort to get the recommended results of Top N.If a unique recommendation result is needed,the calculated click probability is used as the scoring function,and the index mechanism method is used to obtain the recommendation result.Experiments have proved that in the Personal Rank algorithm based on differential privacy,the attacker cannot make difference through multi-account query,and will not get the operation behavior of the target user or other users on the item.Therefore,the algorithm ensures the security of personal information in the recommendation process,and at the same time can provide better recommendation results.Secondly,for different companies with large amounts of data,they usually do not directly share private user data,and there is a lack of information communication mechanisms that can trust each other,resulting in relatively closed information interaction.In order to share data's information between different companies,the key is how to train together to get results without disclosing the original data of the companies.Therefore,consider using homomorphic encryption and differential privacy mechanisms to introduce a third party responsible for issuing homomorphic encryption keys,so that in the federated learning training model,participants can obtain model results without directly using each other's data.By comparing the vertical federated learning with or without fusion of differential privacy,it is found that the accuracy of the training model is not much different,and after the differential privacy is added,there is no obvious negative situation,and the potential danger of third-party leakage of keys is avoided,and the federated learning is guaranteed data security in the process.Thirdly,based on the above two aspects of research,a set of movie recommendation system based on differential privacy has been built to ensure the privacy of users and enterprises while recommending movies to users.The two companies provide data and labels.For users who cannot be matched,a Personal Rank algorithm based on differential privacy is used to obtain recommendation results;users who can match each other,under the vertical federated learning of fusion differential privacy,construct a logistic regression model,And calculate the click-through rate of the target user's movie through a third party,combine the click-through rate obtained by random walk,and weight the new click-through rate,and display the recommended movie result at the same time.