摘要:
推荐系统是建立在海量数据挖掘基础之上的一种智能平台，根据用户个人信息与物品特征，比如用户的兴趣、历史购买行为和物品的材质、价格等，利用统计分析和机器学习等人工智能技术建立模型，预测用户对新物品的评价与喜好，从而向用户推荐其可能感兴趣的潜在物品，以实现个性化的信息服务和决策支持.然而，推荐系统的历史数据集、预测模型和推荐结果都与用户的隐私休戚相关，如何能在有效保护用户隐私的前提下，提供正确性可验证的有效推荐结果是一个具有挑战性的重要研究课题.国内外现有的工作多是通过数据扰动或公钥全同态加密技术来试图解决这个问题，但都无法满足推荐系统对高效性、精确性和各类隐私保护的要求.从推荐系统隐私保护的模式、安全模型、轻量级的推荐系统隐私保护一般性构造与推荐结果正确性可验证、可审计等方面，系统阐述了国内外最新研究成果，并在此基础上提出了存在问题、未来研究方向与解决方案.在安全模型方面，聚焦于标准模型或通用组合模型下，用户数据隐私、预测模型隐私和推荐结果隐私等多种安全模型的形式化刻画；在轻量化方面，将不依赖公钥全同态加密技术，通过减少公钥加密/解密次数(最优时一次)，在单用户、多数据模型和多用户、多数据模型下，提出高效的推荐系统隐私保护一般性构造方法；最后，通过批量验证技术研究推荐结果轻量化防欺诈与抗抵赖的一般性理论问题.从而，为适用于推荐系统隐私保护的新型加密方案研究及其实用化提供理论和方法支撑.
Abstract:
Recommender system is a type of intelligent platform based on massive dataset mining, which can establish recommendation model, predict users’ preferences on unrated items and achieve individualized information service and strategy support by exploiting the techniques of statistic analyzing, machine learning and artificial intelligence, according to the unique profiles of users and the different characteristics of various items, such as users’ interests, historical consumption behaviors, the quality and the prices of items. Unfortunately, the historical dataset, prediction model and recommendation result are closely related to the users’ privacy. How to provide accurate prediction results under the conditions that the users’ privacy is well protected and the correctness of the recommendation result is efficiently verified becomes a challenging issue. The state-of-the-art mainly focused on solving this problem, by using the techniques of data perturbation and public key fully homomorphic encryption (FHE). However, most of them cannot satisfy all the requirements of accuracy, efficiency and types of privacy preserving required by recommender systems. This article elaborates the existing work from the following four aspects, namely the operation mode, formal security model, the generic constructions of lightweight privacy preserving recommender system and the verification, and the accountability of recommendation results; and identifies the unaddressed challenging problems with convincing solutions. For security models, we focus on formalizing the security models with respect to user data privacy, prediction model privacy and recommendation result privacy, under the standard model or universal composable (UC) model. For efficiency, without exploiting public key FHE, we study the generic constructions of efficient privacy preserving recommender system, respectively in the single user, multiple data setting and the multiple user, multiple data setting, by reducing the usage times of public key encryption and decryption (i.e. only once while it is optimized). Last but not least, we also address the generic theoretical issue of efficient correctness verifiability and auditability for recommendation results, by exploiting the technique of batch verification. All the convincing techniques and solutions discussed above would significantly contribute to both the theoretical breakthrough and the practicability for privacy preserving in recommender systems.
百度学术
下载:
