Federated recommender system based on mask protection

In today's era of data expansion, people can obtain a large number of different kinds of information through different channels, but this will lead to the emergence of "information overload". However, as private information becomes more and more accessible, governments around the world pay more and more attention to data privacy, and have introduced various laws and regulations to ensure that data is not allowed to be obtained and disseminated at will, which leads to the phenomenon of "data islands", where formal platforms cannot obtain data for research, and a large amount of valuable data is wasted. Federated learning provides an idea to solve this problem, and the data can be provided without localization, and it can also provide reliable machine learning training. However, the latest research shows that even if the local data is not uploaded, it can be hacked by malicious actors, and that traditional privacy protection methods can lead to lower usability of the data. In order to solve this problem, this paper proposes to use mask protection technology in the state-of-the-art Neural Cooperative Filtering (NCF) network structure to protect data privacy. Due to the different data quality and computing power of each client, the loss of each client is too different. Therefore, this paper designs a loss-based aggregation algorithm to reduce the loss difference of each client training. Experiments show that the system FedMPRec achieves a recommendation quality comparable to that of the original federal recommendation, and the method proposed in this paper can complete privacy protection without compromising data availability.