Research On Recommendation Algorithms Based On Differential Privacy Protection

The continuous development and application of Internet technology has brought unprecedented speed and channels for people to obtain information,but also caused a serious information overload problem.Recommendation system is one of the effective tools to solve the information overload problem,which can analyze and predict users’ needs and preferences based on their historical behaviors,interests and social connections,and then provide personalized recommendations to users.At present,recommendation systems have been widely used in e-commerce,social networks,and entertainment,etc.However,with the increasing application of recommendation systems,the problem of user privacy leakage is becoming more and more serious,and the construction of privacy-preserving recommendation systems is imminent.In this paper,we analyze the application of privacy protection technology on recommendation systems,and select the differential privacy method that can achieve privacy protection without complicated encryption and decryption process and only add noise to solve the data security problem in recommender systems,further investigate the existing differential privacy protection recommendation system solutions and summarize their four main limitations,namely: few studies combine differential privacy techniques with deep learning recommendation algorithms,most studies focus on a single data source and a single recommendation algorithm,only consider the scenario of trusted servers,and do not consider the different distributions of different data sets and different value sensitivity of datasets.Finally,this paper designs two recommendation algorithm models based on differential privacy protection to address the above four limitations,which are summarized as follows:(1)Given the limited research on combining differential privacy techniques with deep learning recommendation algorithms,this paper aims to explore how differential privacy techniques can be applied to deep learning recommendation algorithms in the context of trusted servers.Specifically,this paper designs the DPAuto Rec algorithm based on the Auto Rec algorithm,using the Laplace mechanism,Gaussian mechanism to perturb the data to satisfy the differential privacy and Rényi differential privacy to further optimize the consumption of privacy budget to improve the model accuracy.The simulation experimental results show that DPAuto Rec can still provide valuable prediction results while satisfying differential privacy.(2)In view of the fact that most of the current research only focuses on a single data source and a single recommendation algorithm,and is limited to the scenario of a trusted server,ignoring the different distributions of different datasets and different value sensitivity of datasets,this paper proposes a new privacy protection recommendation system framework,which uses user implicit feedback behavioral data,takes into account the different value sensitivities and data distributions at the client side,and uses LCF-VDP(Local Collaborative Filtering-Value Differential Privacy)mechanism to perturb the original data and upload it to the server;the server mixes the similarity of the two algorithms and finally selects the topk mixed similarity to send to each user device,and performs prediction score calculation and recommendation in each user device.Simulation experimental results show that the proposed method can select appropriate parameters according to different requirements to achieve the best recommendation effect,and LCF-VDP has better utility than the traditional perturbation mechanism under various privacy budgets.