Federated Learning Service Market: A Game Theoretic Analysis

Federated learning enables data owners in intelligent communication systems to share information without revealing actual data contents. In federated learning, each data owner trains data locally, and only uploads the corresponding trained gradient to a learning server. The learning server aggregates collected gradients and further trains a learning model by averaging all the gradients. The trained model can be returned to data owners for improving their performance of data utilization and analysis. In this work, we extend the two-layered federated learning architecture to a learning market with privacy preserving. In the learning market, data owners apply federated learning to trade information extracted from data, a learning service provider collects information from data owners and provides data services based on the learned model to arbitrary service users, who have no data but are willing to pay for data services. To analyze and solve the optimal behaviours of all the participants in the system, a market-oriented architecture is formulated with a Stackelberg game theoretic approach, considering social impacts among all the market participants in the data and model trading processes.