Secure Federated Matrix Factorization via Device-to-Device Model Shuffling

Location-Based Recommendation Systems (LBRS) use device location data to suggest nearby hotels, restaurants, and points of interest. Since directly collecting location data from users can raise privacy concerns, there is growing interest in building recommendation systems based on Federated Learning (FL). Under FL, parameters of recommendation model learned on each user’s device are collected on a single server to build aggregated model. While FL does not raise privacy concerns about data collection since it does not collect user data directly, it may construct unfair models that repeatedly recommend specific locations. Although there are training methods to achieve fair recommendations that prevent such bias, they require more training epochs than usual. In FL, a malicious server can infer the original location data by continuously tracking a specific user’s parameter updates, and the inference accuracy increases proportionally with the number of training epochs. This means that achieving fair location recommendations in FL puts the original data at risk. In this paper, we design a novel parameter aggregation method to build fair and secure FL recommendation models. In the proposed aggregation method, users exchange parameters with each other before model aggregation to prevent malicious servers from inferring the original data. Even if a server (adversary) continuously tracks a specific user’s device, it cannot get parameters from the same user, thus preventing inference of the original location data. An experiment result demonstrated that the proposed method can reduce training time while maintaining the same accuracy as homomorphic encryption approach.