Federated-Learning-Empowered Distribution Training for Generative Artificial Intelligence in Vehicular Networks

Le résumé fourni par la source

Generative artificial intelligence (GAI), e.g., diffusion model is recognized as a promising paradigm for enhancing intelligent transportation systems in vehicular networks. However, the existing implementation of GAI in vehicular networks is limited due to the massive data requirements of GAI and the considerable resources for model training, particularly in distributed vehicular network environments. Federated learning (FL) offers a promising solution by enabling distributed collaborative training for GAI. Therefore, in this paper we present an FL-empowered diffusion model training scheme for vehicular networks. Specifically, first, a novel utility evaluation model based on local model training accuracy is designed to assess the contribution of each vehicle's local model. The interactions between the edge computing servers and vehicles are modeled using a Stackelberg game, while a non-cooperative game determines the optimal strategy among vehicles. To account for the heterogeneity of vehicles and the uncertainty of associated risks, we incorporate prospect theory (PT) to represent subjective utility. Afterward, a backward induction mechanism is devised to determine the Stackelberg equilibrium for deriving the optimal decisions of edge computing servers and vehicles. Finally, simulations are conducted to illustrate that the proposed scheme significantly improves the sum utility rate in comparison to other baseline schemes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.