Preserving User Privacy For Large-Scale Personalized Online Video Service

The large scale online video systems not only occupy the major internet traffic and market share,but also keep a huge advantage in both the number of users and the valid browsing time.Personalization recommendation service has become the principal competitive mean of video websites and brought the leakage risk of user privacy simultaneously.On one side,recommender systems can infer user profile information,e.g.,gender,age,etc.,meanwhile,lead to user privacy leakage.On the other side,the attackers disguising as normal users can obtain user history behavior records directly from the recommended output,and further infer the sensitive interest preference of the target user.This type of non-direct access attacks poses more serious threats to user privacy.In fact,against the contradiction between personalization recommendation and user privacy preservation,the existing privacy-preserving recommendations generally make tradeoffs with the contradiction.And it is a consensus of current research that protecting user privacy will lose the recommendation performance.For the large scale online video systems,whether and how to preserve user privacy information and guarantee or even improve the recommendation service quality simultaneously is the current research focus.Meanwhile,it is also an urgent issue for other large scale online service systems.Targeting the problem described above,this paper analyses the high leakage risk of user privacy in the large scale online video systems firstly,as well as the possibility of accurate user privacy inference with a small amount of user browsing history based on gender information.And then,to protect both user identifiable information,e.g.,gender,age etc.and sensitive interest preferences,the recommendation-friendly privacy-preserving framework and the differential privacy collaborative filtering algorithm based on topic privacy-relevance are proposed respectively in this paper,which both protect user privacy and guarantee or even improve the recommendation service quality simultaneously.The main works and innovations of this paper are as follows.Firstly,in terms of user privacy inference,to solve the high data sparseness problem in the actual online video systems,two different user behavior aggregation methods towards Chinese and English video systems are proposed respectively.Specifically,for Asian language with no separator,a simple and effective keyword extraction algorithm is designed.And for English video system,a new user behavior aggregation method is proposed based on synonyms with almost no loss of information.To solve the gender imbalance distribution problem,this paper improve the privacy inferring model based on the new evaluation metric.The experiments from several large scale online video systems show that,compared with the existing work,the methods proposed in this paper not only solve the high data sparseness and gender imbalance problem,but also make the gender inference result achieve global optimal.Furthermore,the study verifies the possibility of exposing user privacy on a small amount of data records in a video system with high data sparseness.Secondly,in order to protect user privacy information like age and gender without losing recommendation performance,the recommendation-friendly privacy-preserving framework is proposed in this paper.The existing methods add factitious ratings of videos favored by the opposite user group,sacrificing the recommendation accuracy to obscure user information.This approach overlooks an important fact that an individual user might like the most popular content in the opposite user group,e.g.,the opposite gender or age group,in a statistical sense.With this observation,this paper proposes a new video-similarity computing method,and designs the video selection strategies to obscure user gender(age)information and strengthen interest preferences.And then,the factitious rating estimation method is introduced.Experiments show that,compared with the tradeoff approaches,the recommendation-friendly privacy-preserving framework proposed in this paper can both protect user information,e.g.,gender,age etc.and guarantee or even improve the recommendation service quality simultaneously.Besides,the framework can be generalized to book,CD and music recommendation systems as well.Thirdly,for the typical non-direct access attacks towards user behavior records,the differential privacy collaborative filtering algorithm based on topic privacy-relevance is proposed in this paper.The existing differential privacy collaborative filtering algorithms provide same protection effort on different user behavior records.Although the mean absolute error of recommendation is acceptable,the performance of Top-k recommendation which is widely used in actual system,does decrease seriously.To this point,combining both the observation that user sensitivity is different to the disclosure of different behavior records and the consideration of the sparseness of user behavior in video system,the work in this paper protects user privacy based on topic privacy-relevance,and provides stronger protection for high privacy-relevance topic under the equal privacy budget.To improve the quality of personalization recommendation service,the recommended output is reordered and filtered on the user client to recommend the final Top-k videos based on user interest preference.Experiments prove that,under the equal privacy budget,the differential privacy collaborative filtering algorithm based on topic privacy-relevance can not only protect user interest preferences differently,but also improve the precision and recall of the Top-k recommended videos in collaborative filtering systems simultaneously.