Heterogeneous Federated Learning for Vehicle-to-Everything: Feature Prototype Aggregation and Generative Feedback Mechanism

With the rapid advancement of Vehicle-to-Everything (V2X) technology, there is a growing demand for collaborative perception among vehicles and multimodal devices (e.g., roadside units, pedestrian terminals). However, traditional centralized learning and federated learning (FL) face challenges in model convergence and performance degradation due to non-IID data distribution, privacy protection requirements, and communication bandwidth constraints among massive heterogeneous devices in V2X scenarios. To address these issues, this paper proposes a heterogeneous federated learning framework based on feature prototype alignment and generative knowledge transfer, enabling efficient and secure cross-device collaborative learning. The framework employs dynamic edge-enhanced contrastive learning on the server side to generate trainable global feature prototypes. These prototypes are subsequently decoded into composite images through a pre-trained generative adversarial network, achieving lightweight privacy-preserving knowledge transfer. Experimental results on CIFAR-10, CIFAR-100, and BelgiumTSC datasets demonstrate that our method achieves significant accuracy improvements compared with baseline approaches such as FedDistill and FedTGP. This study establishes a novel theoretical framework and technical pathway for collaborative learning in V2X environments that effectively balances privacy protection with model performance.