Abstract
The emergence of intelligent vehicular networks has catalyzed the advancement of communication architectures capable of meeting the rigorous demands of ultra-reliable low-latency communication (URLLC), enhanced spectral efficiency, and interference mitigation. This paper presents a novel cross-layer interference management framework for Vehicle Road Cooperation Systems (VRCS), built on the integration of Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surfaces (STAR-RIS) with Generative Artificial Intelligence (GAI)-enhanced Tiny Federated Learning (TinyFL). High mobility and dynamic topologies in VRCS often result in severe co-channel interference, adversely affecting link reliability and latency. To address this, the proposed system leverages the full-space beamforming capabilities of STAR-RIS to intelligently reconfigure the wireless propagation environment. Complementing this, a lightweight and privacy-preserving TinyFL architecture is employed, tailored to the computational constraints of vehicular nodes. To further improve model generalization and adaptability under non-IID and time-varying scenarios, GAI is integrated to synthesize diverse and realistic training data. The framework supports decentralized and intelligent mode selection, link scheduling, and resource allocation through a Markov Decision Process (MDP)-driven reinforcement learning formulation spanning the physical, MAC, and network layers. This design ensures scalable, autonomous decision-making while preserving data privacy. Extensive simulations demonstrate the superiority of the proposed system, achieving up to 32.4% improvement in interference suppression, 27.8% gain in throughput, and 41.6% faster learning convergence compared to conventional FL, FRL, and TinyFRL baselines. These results highlight the effectiveness of combining STAR-RIS with edge AI to address the critical challenges of scalability, latency, and adaptability in next-generation vehicular communication systems.
Funding Information
  • Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia, through the Researchers Supporting Project (PNURSP2025R333)