Secure Federated Matrix Factorization via Shuffling Encrypted Parameters Between Devices

Recommendation systems that provide clients with suggested locations, such as hotels, restaurants, and geographical objects related to their current location, based on location data measurable from devices like GPS, are becoming increasingly valuable. As collecting location data directly from clients can lead to privacy issues, there has recently been a growing expectation to build recommendation systems using Federated Learning (FL). In FL, instead of collecting data from clients, the parameters of the recommendation model trained on each client’s device are aggregated on a server to build a single model. Since FL does not collect data from customers, there is no concern about privacy leakage, but recommendation models built with FL are subject to bias during training. To prevent biased recommendations, there are training methods that guarantee fair recommendations. However, these methods require extensive training epoch, and if the malicious server continuously tracks the parameters, the original location can be inferred from the update rate in the parameters. In this study, we designs a novel parameter aggregation in which clients exchange parameters with each other before server aggregation, preventing malicious servers from guessing the original location data. Even if a malicious server continues to trace a particular client’s device, it cannot receive same client’s parameters, thereby preventing the inference of original location data.