Improved BCoop, a Cooperation-Promoting Reward Scheme in Heterogeneous Federated Learning

In the era of big data, update of models increasingly rely on federated learning, the synthesis of training results from distributed nodes. Ensuring equitable reward distribution is essential to maintaining active participation and effective contribution in each phase of federated learning such as training and data transmission. We propose in this paper a new reward scheme tailored for nodes in a federated learning network that contribute and relay data. Our approach centers on a heterogeneous setting, contrasting with the commonly explored homogeneous scenarios in existing literature. In our model, nodes have the discretion to participate in various networks, reflecting the competitive and diverse nature of industrial applications. As it is inappropriate to blindly replicate the results in homogeneous setting to heterogeneous setting (e.g. a node might not have 100% probability of participating into our network due to existence of alternatives), our exploration of reward schemes in heterogeneous context has originality and ingenuity. There are other advantages of our scheme including 1). Simplicity: it is designed for a two-player scenario with one player having a predetermined likelihood of non-participation in our network; 2). Efficiency: it necessitates only a single communication exchange, minimizing potential misunderstandings, and 3). Collaborative incentivization: it fosters cooperation under the existence of “honest node”, and guarantees transmission of both player’s messages under the existence of “participative node” (both terms are defined in section IV). Overall, our scheme not only introduces a realistic approach to reward distribution in heterogeneous federated learning environments but also promotes cooperation and effective communication among participating nodes.