一种基于差分隐私保护的协同过滤推荐方法 A Collaborative Filtering Recommendation Method Based on Differential Privacy

2、一种基于差分隐私保护的协同过滤推荐方法
A Collaborative Filtering Recommendation Method Based on Differential Privacy

摘要：由于推荐系统需要利用大量用户数据进行协同过滤，会给用户的个人隐私带来相当大的风险，如何保护隐私数据成为推荐系统当前面临的重大挑战.差分隐私作为一种新出现的隐私保护框架，能够防止攻击者拥有任意背景知识下的攻击并提供有力的保护.针对推荐系统中的隐私保护问题，提出一种满足差分隐私保护的协同过滤推荐算法.首先，构建用户和项目的潜在特征矩阵，有效降低数据稀疏性；然后，采用目标扰动方法对矩阵中添加满足差分隐私约束的噪声得到噪矩阵分解模型；通过随机梯度下降算法最小化相关联的正则化平方误差函数来获取模型中的参数；最后，应用差分隐私矩阵分解模型进行评分预测，并在MovieLens和Netflix数据集上对算法的有效性进行评价.实验结果证明:所提出方法的有效性能够在有限的精度损失范围内进行推荐并保护用户隐私.

关键词: 差分隐私, 隐私保护, 协同过滤, 推荐系统, 矩阵分解

Abstract：Collaborative filtering with large amount of user data will raise serious risk privacy of individuals. How to protect private data information from disclosure has become one of the greatest challenges to recommender systems. Differential privacy has emerged as a new paradigm for privacy protection with strong privacy guarantees against adversaries with arbitrary background knowledge. Although several studies explored privacy-enhanced neighborhood-based recommendations, little attention has been paid to privacy preserving latent factor models. To address the problem of privacy preserving in recommendation systems, a new collaborative filtering recommendation algorithm based on differential privacy is proposed in this paper, which achieves trade-off between recommendation accuracy and privacy by matrix factorization technique. Firstly, user and item latent feature matrices are constructed for decreasing sparsity. After that, matrix factorization model with noise is generated by adding the differential noisy using objective perturbation method, and then stochastic gradient descent is utilized to minimize regularized squared error function and learn the parameters of model. Finally, we apply a differentially private matrix factorization model to predict the ratings and conduct experiments on the MovieLens and Netflix datasets to evaluate its effectiveness. The experimental results demonstrate that our proposal is efficient and has limited side effects on the precision of recommendation.

Key words: differential privacy, privacy protection, collaborative filtering, recommender systems, matrix factorization

collaborative：adj合作的
privacy：n.隐私
filter：v.过滤
disclosure：v.公开，泄露
paradigm：n.范例
adversary：n.对手，敌人
latent：adj.潜在的
algorithm：n.算法
matrix：n.矩阵
factorization：n.分解
decrease：v.降低
sparsity：n.稀疏
generate：v.产生
perturbation：n.扰乱
stochastic：adj.随机的
gradient：n.梯度
utilize：v,利用
squared error：n.平方误差
function：n.函数
parameter：n.参数
experiment：n.实验
demonstrate：v.证明
precision：n.精确度