A Privacy-Preserving Oriented Service Recommendation Approach based on Personal Data Cloud and Federated Learning

Abstract

Personal data cloud, as an emerging personal data management mode in recent years, enables to reduce the risk of privacy disclosure and protect the rights and interests of individuals given by privacy protection laws and regulations. Personal data that is generated during the interaction between individual users and various services contains a lot of useful personalized but private information and plays a crucial part in personalized service recommendation. In traditional service recommendation scenario, personal data of massive users is centralized owned/managed by service providers, which is easy to lead to privacy disclosure and personal data abuse. In the personal data cloud based service recommendation scenario, personal data of individual users is distributed stored and controlled by users themselves. To address the challenges of privacy protection and distributed storage of personalized data in this new recommendation scenario, we propose HyFL, a deep learning based recommendation algorithm with hybrid federated learning. HyFL can conduct recommendation based on the personal data from multiple services. The security of HyFL is theoretically proved, and experiments on real-world datasets demonstrate that HyFL performs better on the basis of privacy preservation than that of some traditional recommendation approaches.