Privacy Recommendation Algorithm Based On Sequential Item2vec

In the era of big data,recommendation algorithms are far and wide used in many fields.In order to avoid information overload,the role of the algorithms becomes more and more important.Recommendation systems usually use data such as userapplication integration records and users’ attributes to filter the information,but incidents are that data get leaked and users’ private information is exposed,which would cause problems like malicious recommendations.Combining the recommendation algorithm with the privacy preserving mechanism can better protect users’ privacy while providing users with accurate recommendation information.This thesis proposes a recommendation algorithm named Priv Item2 Vec,to protect users’ privacy.First,this thesis introduces the bi LSTM algorithm into item2 vec recommendation.Bi LSTM can make the item2 vec algorithm time-sequential and also make the accuracy of result higher.Second,this thesis uses Paillier algorithm to encrypt and reconstruct item id and customer’s behavior sequence,and it makes the protection of user’s privacy effectively.The Priv Item2 Vec algorithm first inputs the item sequence into the item2 vec model to obtain a static vector.Then it inputs the vector to bi LSTM model for bi-directional training.Then it obtains a sequential result and the Paillier encryption algorithm is used to reconstruct the item id and vector output by item2 vec and bi LSTM to protect the users’ privacy.Finally,the result is sending to the server side.According to the homomorphism of encryption algorithms,a random number is generated to be calculated with the item id on the user side and the result is encrypted to construct a pseudo item id,which will then be sent to the server side together with the encrypted random number.When data gets to the server side,the real encrypted item id is also calculated based on the homomorphic encryption algorithm,and the right product is found from the mapping sequence in the server.Then the sever uses cosine similarity method to calculate the similarity among vectors,and finally heap sorting algorithm will sort similarity sequences,obtaining the Top-K recommended sequence and sending it to the user side.User side decrypts the sequence to obtain a real result.In the recommended calculation process,only the user side retains the secret key of the encryption algorithm,so even the server side cannot decrypt it.In order to show and prove the performance of Priv Item2 Vec,experiments were ran on the real data set Movie Lens.And the results got compared with BNR,NCF and item2vec-based which were commonly used recently.Priv Item2Vec’s recommended hit rate,n DCG result,precision and other indicators are all better than other algorithms.When comparing the calculation efficiency before and after encryption,the average processing time difference on the server side is 0.148 s,which confirms that the time consumed by the Priv Item2 Vec algorithm in the actual application process has almost no effect on the overall running time.The Priv Item2 Vec algorithm based on sequential item2 vec makes the traditional item2 vec algorithm sequential by combining with bi LSTM and improves the accuracy.Meanwhile,the use of Paillier encryption algorithm ensures the data obtained by sever side is all encrypted or embedding vectors to reduce the readability of the data.By using this algorithm,we can prevent attack from semi-honest servers,avoid malicious recommendations caused by server-side data leakage,and protect transmission between client and server.The future work will be exploring the influence of integrating social network factors into the privacy preserving recommendation algorithm.Because the recommendation results can be affected by preferences of friends with different degrees of intimacy.Meanwhile the future work will focus on how to defend against semi-honest or malicious friends in social networks.