Distributed Mechanism Design in Continuous Space for Federated Learning Over Vehicular Networks

Abstract

With the rapid development of the 5G network and the promising of 6G network, vehicles can report a large amount of real-time traffic information to Road-Side Units (RSUs). However, due to the large communication cost, limited computing resources, and privacy leakage risks, centralized data processing by RSUs is not efficient and not secure. Thereby, federated learning is introduced to enable vehicles to train local models and send the model parameters to the RSUs, without the need for revealing their personal data. Nevertheless, without an efficient incentive mechanism, the vehicles, as data owners, may be unwilling to join the federated learning task. In this paper, we adopt contract theory to design an incentive mechanism with asymmetric information and continuum of types for interaction between the RSU and vehicles. The designed mechanism satisfies the incentive compatibility, individual rationality and also maximizes the RSU's profit. The technical challenge is to solve a non-quadratic functional optimization in continuous space with coupling constraints among the vehicles. To address this challenge, we propose two iterative algorithms to find the RSU's optimal strategies as the functions of vehicles' types. In the first algorithm, a combination of the calculus of variation and dual decomposition methods is utilized to achieve an analytical solution for the optimal strategy of RSU, while in the second one, a combination of approximate dynamic programming and neural networks is used to estimate the optimal strategy of RSU.