基于隐私保护的联邦推荐算法综述

推荐系统通过集中式的存储与训练用户对物品的海量行为信息以及内容特征, 旨在为用户提供个性化的信息服务与决策支持. 然而, 海量数据背后存在大量的用户个人信息以及敏感数据, 因此如何在保证用户隐私与数据安全的前提下分析用户行为模式成为了近年来研究的热点. 联邦学习作为新兴的隐私保护范式, 能够协调多个参与方通过模型参数或者梯度等信息共同学习无损的全局共享模型, 同时保证所有的原始数据保存在用户的终端设备, 较之于传统的集中式存储与训练模式, 实现了从根源上保护用户隐私的目的, 因此得到了众多推荐系统领域研究学者们的广泛关注. 基于此, 对近年来基于联邦学习范式的隐私保护推荐算法进行全面综述、系统分类与深度分析. 具体的, 首先综述经典的推荐算法以及所面临的问题, 然后介绍基于隐私保护的推荐系统与目前存在的挑战, 随后从多个维度综述结合联邦学习技术的推荐算法, 最后对该方向做出系统性的总结并对未来研究方向与发展趋势进行展望.

A Survey on Privacy-preserving Federated Recommender Systems

Recommender systems aim to provide users with personalized information services and decision support, by mining the centrally stored historical behaviors of users on items and their inherent attributes. However, there is numerous users＇sensitive information behind the massive data, therefore, how to mine users＇behavior patterns on the premise of ensuring users＇privacy and data security has become a hotspot. As an emerging privacy-preserving paradigm, federated learning can coordinate multiple participants to jointly train a lossless global shared model by transmitting model parameters or gradients, and ensure that all original data are saved in the local terminal devices, compared with the traditional centralized storage and training mode. Therefore, it has been widely concerned by many researchers in the field of recommender systems. Based on this, the paper will provide a comprehensive survey, systematic taxonomy, and in-depth analysis of federated recommender systems. Specifically, we first summarize the classical recommendation algorithms and potential problems; then introduce the taxonomy and current challenges on privacy-preserving recommendation methods, and then summarize the privacy-preserving federated recommendation models from multiple dimensions. Finally, we conclude this paper and discuss possible research directions in this area.