Incentivizing Data Contributions in DecentralizedCollaborative Learning

Abstract

In a collaborative learning scheme such as the federated learning model, each user benefits from the data contribution of others. Previous work shows that the federated learning protocol can incentivize users to contribute more than in the competitive equilibrium by penalizing deviations. However, a central controller with access to all the data may raise privacy concerns. In this work, we construct a decentralized collaborative protocol in which users share data without relying on a centralized controller. We then extend this protocol to a repeated game and analyze the competitive equilibrium behavior, along with strategies users can implement to foster collaboration in the repeated setting of the protocol. We provide a quantitative analysis of free-rider behavior under decentralized protocols and compare the amount of information collected with decentralized protocols against that in the centralized protocol.