Research On Hybrid Privacy Models And Algorithms For Collaborative Filtering

With the advent of the Internet economy era, recommended system has gradually penetrated into people’s daily lives. A sharp increase in network data so that the user (consumer) are often difficult to find the most suited to their needs or information. With a reliable recommendation system means being able to provide more convenient services to attract users. The collaborative filtering(CF) recommendation system is a mainstream mature recommendation algorithm.CF can be applied to many types of sites. This algorithm to calculate the similarity between two users according to their rating of the items in the past. Then, a recommendation based on rating records rated by similar user of the target user.This recommended approach has many advantages, it has been widely applied to the current personalized recommendation service, but since the recommendation is based on obtain detailed personal information, personal privacy issues therefore be more serious challenge.Currently, the data released research on privacy issues include two of privacy protection system: k-anonymity privacy protection system and differential privacy. Some studies also has applied the k-Anonymous system or differential privacy to the CF each to protect personal information.。 On the existing research results, because the thought of a simple model of k-anonymity and its less complex operations on data, apply k-anonymity alone with the CF made the system relatively easy to implement, but experiments show it is difficult to ensure the usefulness of the generalized data as each tag contains too many items in e-commerce data; Both the recommended privacy and accuracy of the new method to apply differential privacy to CF have good guaranteed, but once the recommended system is under the dynamic environment, the available privacy budget of each data release will be reduced with the increasing total release times of data according to the sequential composition of differential privacy, it’s hard to control the noise.On these privacy issues of CF, to sum up the research of this paper as follows:Firstly, we introduced the present situation of recommendation system. We research for the main process and implementation of CF recommendation algorithm, get its advantages and disadvantages, and find that the major challenge of recommendation system development is the security of user information.Secondly, we analyzed the privacy issues faced by CF. We described the principles and methods of information disclosure which recommendation system faced such as server collects, employee disclosure, residual data, recommended disclosure and so on. We described this type of privacy and security issues as a attack mode which include target, background knowledge, attack methods and the definition of privacy disclosure. Then we introduced two main methods about anonymous protection:k-anonymity and Differential Privacy. We analyzed its application and shortcomings in recommendation system, set up a standard of how to evaluate the privacy model based on the utility and privacy of data.moreover, based on CF techniques’privacy issues, this paper analyzed the existing security and proposed a security model about privacy issues. According to the current security model about privacy issues and the requirements in the practical application of the recommended system for accuracy and privacy, combining two anonymous k-anonymity method and the differential privacy method, this paper proposed a model of ρ-hybrid anonymity recommend model. According to the model, the paper designed the corresponding p-hybrid anonymity algorithm. This algorithm add the limitation of project evaluation diversity on the choice of target user’s k-nearest neighbor, and weighted recommend using the neighborhood similarity of Laplace noise. This algorithm can effectively resist the KNN attacks and maintain good recommendation accuracy and privacy in the released dynamic environment.Finally, based on the proposed model and the improved algorithm combining several the recommended requirements in practical recommendations result, this paper implements a privacy protection recommendations system which adapts to release several times. Comparing several real data sets, the experiments verify the effectiveness of the privacy protection method, and improve the practicability of the data while ensure the data privacy protection at the same time...