Differentially Private Social Recommender Systems

With the rapid development of mobile network and e-commerce,recommender systems play an increasingly important role in our daily life.Recommender systems can alleviate the problem of information overload and help users to make decisions.In recent years,with the continuous expansion of social network,many recommender systems begin to use the social information between users as auxiliary information to optimize their algorithms.These recommender systems are also called social recom-mendation system.Social recommender systems give a more realistic simulation about users' behavior,and can alleviate the problems of cold-start and data sparse.In recent years,with the improvement of regulations and the increasement of citi-zens' privacy awareness,individuals,enterprises,and governments have paid more and more attention to data privacy.Recommender system inevitably face the problem of data privacy as they utilize a lot of user data.Solving the problem of user data pri-vacy in recommender systems is not only to protect the privacy of users,but also to benefit service providers in terms of service quality improvement and users' loyalty in-creasement.However,existing social recommender systems are not perfect to protect the user's privacy,which may lead to user's unwillingness to provide personal ratings,even deliberately give untrue ratings to protect their privacy.Ultimately,these behavior affect the performance of recommender systems.This thesis is based on the above background,and the content is divided into two parts:In the first part,we propose a differentially private social recommender system to solve the privacy problem.First,we point out the privacy problem faced by so-cial recommender systems,the assumption of the adversary,and the privacy goal and accuracy goal of our system.Then,based on three widely used social recommender algorithms which utilizing matrix factorization,combined with the appropriate differ-ential privacy mechanism,we reduce the privacy budget and strenth the protection of user data.Through the experimental evaluation on two real datasets,compared with the existing schemes,our algorithms are verified that they can protect users' ratings,while retaining the accuracy of prediction.In the second part,as the recommender system has more sufficient user data,the prediction results are usually more accurate.In real life,the data of a user may be dis-tributed in different parties.For the sake of interests and regulations,data parties cannot directly exchange the original user data they owned,and they need to use recommender algorithms on vertically partitioned data to cooperate in order to get better recommen-dation performance.Even so,like recommender systems having centralized data,the problem of privacy cannot be avoided.Therefore,we propose a differentially private social recommender system on vertically partitioned data which based on the concept of federated recommendation.To solve the problem of gradient aggregation in federated recommendation,we introduce secure multi-party computation to replace the parameter server in existing schemes,which strengthens the data security.Finally,compared with other algorithms,the efficiency of our algorithm is verified.