Dynamic Games for Social Model Training Service Market via Federated Learning Approach

In recent years, an increasing amount of new social applications have been emerging and developing with the profound success of deep learning technologies, which have been significantly reshaping our daily life, e.g., interactive games and virtual reality. Deep learning applications are generally driven by a huge amount of training samples collected from the users' participation, e.g., smartphones and watches. However, the users' data privacy and security issues have been one of the main restrictions for a broader distribution of these applications. In order to preserve privacy while utilizing deep learning applications, federated learning becomes one of the most promising solutions, which gains growing attention from both academia and industry. It can provide high-quality model training by distributing the training tasks to individual users, relying on on-device local data. To this end, we model the users' participation in social model training as a training service market. The market consists of model owners (MOs) as consumers (e.g., social applications) who purchase the training service and a large number of mobile device groups (MDGs) as service providers who contribute local data in federated learning. A two-layer hierarchical dynamic game is formulated to analyze the dynamics of this market. The service selection processes of MOs are modeled as a lower level evolutionary game, while the pricing strategies of MDGs are modeled as a higher level differential game. The uniqueness and stability of the equilibrium are analyzed theoretically and verified via extensive numerical evaluations.